250c2df04a92c5feee85c6e07c75940dc0af5392
[sfrench/cifs-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  64-bit file support on 64-bit platforms by Jakub Jelinek
16  *      (jj@sunsite.ms.mff.cuni.cz)
17  *
18  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
19  */
20
21 #include <linux/fs.h>
22 #include <linux/time.h>
23 #include <linux/highuid.h>
24 #include <linux/pagemap.h>
25 #include <linux/dax.h>
26 #include <linux/quotaops.h>
27 #include <linux/string.h>
28 #include <linux/buffer_head.h>
29 #include <linux/writeback.h>
30 #include <linux/pagevec.h>
31 #include <linux/mpage.h>
32 #include <linux/namei.h>
33 #include <linux/uio.h>
34 #include <linux/bio.h>
35 #include <linux/workqueue.h>
36 #include <linux/kernel.h>
37 #include <linux/printk.h>
38 #include <linux/slab.h>
39 #include <linux/bitops.h>
40
41 #include "ext4_jbd2.h"
42 #include "xattr.h"
43 #include "acl.h"
44 #include "truncate.h"
45
46 #include <trace/events/ext4.h>
47
48 #define MPAGE_DA_EXTENT_TAIL 0x01
49
50 static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
51                               struct ext4_inode_info *ei)
52 {
53         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
54         __u16 csum_lo;
55         __u16 csum_hi = 0;
56         __u32 csum;
57
58         csum_lo = le16_to_cpu(raw->i_checksum_lo);
59         raw->i_checksum_lo = 0;
60         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
61             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
62                 csum_hi = le16_to_cpu(raw->i_checksum_hi);
63                 raw->i_checksum_hi = 0;
64         }
65
66         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw,
67                            EXT4_INODE_SIZE(inode->i_sb));
68
69         raw->i_checksum_lo = cpu_to_le16(csum_lo);
70         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
71             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
72                 raw->i_checksum_hi = cpu_to_le16(csum_hi);
73
74         return csum;
75 }
76
77 static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
78                                   struct ext4_inode_info *ei)
79 {
80         __u32 provided, calculated;
81
82         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
83             cpu_to_le32(EXT4_OS_LINUX) ||
84             !ext4_has_metadata_csum(inode->i_sb))
85                 return 1;
86
87         provided = le16_to_cpu(raw->i_checksum_lo);
88         calculated = ext4_inode_csum(inode, raw, ei);
89         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
90             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
91                 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
92         else
93                 calculated &= 0xFFFF;
94
95         return provided == calculated;
96 }
97
98 static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
99                                 struct ext4_inode_info *ei)
100 {
101         __u32 csum;
102
103         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
104             cpu_to_le32(EXT4_OS_LINUX) ||
105             !ext4_has_metadata_csum(inode->i_sb))
106                 return;
107
108         csum = ext4_inode_csum(inode, raw, ei);
109         raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
110         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
111             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
112                 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
113 }
114
115 static inline int ext4_begin_ordered_truncate(struct inode *inode,
116                                               loff_t new_size)
117 {
118         trace_ext4_begin_ordered_truncate(inode, new_size);
119         /*
120          * If jinode is zero, then we never opened the file for
121          * writing, so there's no need to call
122          * jbd2_journal_begin_ordered_truncate() since there's no
123          * outstanding writes we need to flush.
124          */
125         if (!EXT4_I(inode)->jinode)
126                 return 0;
127         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
128                                                    EXT4_I(inode)->jinode,
129                                                    new_size);
130 }
131
132 static void ext4_invalidatepage(struct page *page, unsigned int offset,
133                                 unsigned int length);
134 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
135 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
136 static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
137                                   int pextents);
138
139 /*
140  * Test whether an inode is a fast symlink.
141  */
142 int ext4_inode_is_fast_symlink(struct inode *inode)
143 {
144         int ea_blocks = EXT4_I(inode)->i_file_acl ?
145                 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
146
147         if (ext4_has_inline_data(inode))
148                 return 0;
149
150         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
151 }
152
153 /*
154  * Restart the transaction associated with *handle.  This does a commit,
155  * so before we call here everything must be consistently dirtied against
156  * this transaction.
157  */
158 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
159                                  int nblocks)
160 {
161         int ret;
162
163         /*
164          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
165          * moment, get_block can be called only for blocks inside i_size since
166          * page cache has been already dropped and writes are blocked by
167          * i_mutex. So we can safely drop the i_data_sem here.
168          */
169         BUG_ON(EXT4_JOURNAL(inode) == NULL);
170         jbd_debug(2, "restarting handle %p\n", handle);
171         up_write(&EXT4_I(inode)->i_data_sem);
172         ret = ext4_journal_restart(handle, nblocks);
173         down_write(&EXT4_I(inode)->i_data_sem);
174         ext4_discard_preallocations(inode);
175
176         return ret;
177 }
178
179 /*
180  * Called at the last iput() if i_nlink is zero.
181  */
182 void ext4_evict_inode(struct inode *inode)
183 {
184         handle_t *handle;
185         int err;
186
187         trace_ext4_evict_inode(inode);
188
189         if (inode->i_nlink) {
190                 /*
191                  * When journalling data dirty buffers are tracked only in the
192                  * journal. So although mm thinks everything is clean and
193                  * ready for reaping the inode might still have some pages to
194                  * write in the running transaction or waiting to be
195                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
196                  * (via truncate_inode_pages()) to discard these buffers can
197                  * cause data loss. Also even if we did not discard these
198                  * buffers, we would have no way to find them after the inode
199                  * is reaped and thus user could see stale data if he tries to
200                  * read them before the transaction is checkpointed. So be
201                  * careful and force everything to disk here... We use
202                  * ei->i_datasync_tid to store the newest transaction
203                  * containing inode's data.
204                  *
205                  * Note that directories do not have this problem because they
206                  * don't use page cache.
207                  */
208                 if (ext4_should_journal_data(inode) &&
209                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
210                     inode->i_ino != EXT4_JOURNAL_INO) {
211                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
212                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
213
214                         jbd2_complete_transaction(journal, commit_tid);
215                         filemap_write_and_wait(&inode->i_data);
216                 }
217                 truncate_inode_pages_final(&inode->i_data);
218
219                 goto no_delete;
220         }
221
222         if (is_bad_inode(inode))
223                 goto no_delete;
224         dquot_initialize(inode);
225
226         if (ext4_should_order_data(inode))
227                 ext4_begin_ordered_truncate(inode, 0);
228         truncate_inode_pages_final(&inode->i_data);
229
230         /*
231          * Protect us against freezing - iput() caller didn't have to have any
232          * protection against it
233          */
234         sb_start_intwrite(inode->i_sb);
235         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
236                                     ext4_blocks_for_truncate(inode)+3);
237         if (IS_ERR(handle)) {
238                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
239                 /*
240                  * If we're going to skip the normal cleanup, we still need to
241                  * make sure that the in-core orphan linked list is properly
242                  * cleaned up.
243                  */
244                 ext4_orphan_del(NULL, inode);
245                 sb_end_intwrite(inode->i_sb);
246                 goto no_delete;
247         }
248
249         if (IS_SYNC(inode))
250                 ext4_handle_sync(handle);
251         inode->i_size = 0;
252         err = ext4_mark_inode_dirty(handle, inode);
253         if (err) {
254                 ext4_warning(inode->i_sb,
255                              "couldn't mark inode dirty (err %d)", err);
256                 goto stop_handle;
257         }
258         if (inode->i_blocks)
259                 ext4_truncate(inode);
260
261         /*
262          * ext4_ext_truncate() doesn't reserve any slop when it
263          * restarts journal transactions; therefore there may not be
264          * enough credits left in the handle to remove the inode from
265          * the orphan list and set the dtime field.
266          */
267         if (!ext4_handle_has_enough_credits(handle, 3)) {
268                 err = ext4_journal_extend(handle, 3);
269                 if (err > 0)
270                         err = ext4_journal_restart(handle, 3);
271                 if (err != 0) {
272                         ext4_warning(inode->i_sb,
273                                      "couldn't extend journal (err %d)", err);
274                 stop_handle:
275                         ext4_journal_stop(handle);
276                         ext4_orphan_del(NULL, inode);
277                         sb_end_intwrite(inode->i_sb);
278                         goto no_delete;
279                 }
280         }
281
282         /*
283          * Kill off the orphan record which ext4_truncate created.
284          * AKPM: I think this can be inside the above `if'.
285          * Note that ext4_orphan_del() has to be able to cope with the
286          * deletion of a non-existent orphan - this is because we don't
287          * know if ext4_truncate() actually created an orphan record.
288          * (Well, we could do this if we need to, but heck - it works)
289          */
290         ext4_orphan_del(handle, inode);
291         EXT4_I(inode)->i_dtime  = get_seconds();
292
293         /*
294          * One subtle ordering requirement: if anything has gone wrong
295          * (transaction abort, IO errors, whatever), then we can still
296          * do these next steps (the fs will already have been marked as
297          * having errors), but we can't free the inode if the mark_dirty
298          * fails.
299          */
300         if (ext4_mark_inode_dirty(handle, inode))
301                 /* If that failed, just do the required in-core inode clear. */
302                 ext4_clear_inode(inode);
303         else
304                 ext4_free_inode(handle, inode);
305         ext4_journal_stop(handle);
306         sb_end_intwrite(inode->i_sb);
307         return;
308 no_delete:
309         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
310 }
311
312 #ifdef CONFIG_QUOTA
313 qsize_t *ext4_get_reserved_space(struct inode *inode)
314 {
315         return &EXT4_I(inode)->i_reserved_quota;
316 }
317 #endif
318
319 /*
320  * Called with i_data_sem down, which is important since we can call
321  * ext4_discard_preallocations() from here.
322  */
323 void ext4_da_update_reserve_space(struct inode *inode,
324                                         int used, int quota_claim)
325 {
326         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
327         struct ext4_inode_info *ei = EXT4_I(inode);
328
329         spin_lock(&ei->i_block_reservation_lock);
330         trace_ext4_da_update_reserve_space(inode, used, quota_claim);
331         if (unlikely(used > ei->i_reserved_data_blocks)) {
332                 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
333                          "with only %d reserved data blocks",
334                          __func__, inode->i_ino, used,
335                          ei->i_reserved_data_blocks);
336                 WARN_ON(1);
337                 used = ei->i_reserved_data_blocks;
338         }
339
340         /* Update per-inode reservations */
341         ei->i_reserved_data_blocks -= used;
342         percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
343
344         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
345
346         /* Update quota subsystem for data blocks */
347         if (quota_claim)
348                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
349         else {
350                 /*
351                  * We did fallocate with an offset that is already delayed
352                  * allocated. So on delayed allocated writeback we should
353                  * not re-claim the quota for fallocated blocks.
354                  */
355                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
356         }
357
358         /*
359          * If we have done all the pending block allocations and if
360          * there aren't any writers on the inode, we can discard the
361          * inode's preallocations.
362          */
363         if ((ei->i_reserved_data_blocks == 0) &&
364             (atomic_read(&inode->i_writecount) == 0))
365                 ext4_discard_preallocations(inode);
366 }
367
368 static int __check_block_validity(struct inode *inode, const char *func,
369                                 unsigned int line,
370                                 struct ext4_map_blocks *map)
371 {
372         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
373                                    map->m_len)) {
374                 ext4_error_inode(inode, func, line, map->m_pblk,
375                                  "lblock %lu mapped to illegal pblock "
376                                  "(length %d)", (unsigned long) map->m_lblk,
377                                  map->m_len);
378                 return -EFSCORRUPTED;
379         }
380         return 0;
381 }
382
383 int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
384                        ext4_lblk_t len)
385 {
386         int ret;
387
388         if (ext4_encrypted_inode(inode))
389                 return ext4_encrypted_zeroout(inode, lblk, pblk, len);
390
391         ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
392         if (ret > 0)
393                 ret = 0;
394
395         return ret;
396 }
397
398 #define check_block_validity(inode, map)        \
399         __check_block_validity((inode), __func__, __LINE__, (map))
400
401 #ifdef ES_AGGRESSIVE_TEST
402 static void ext4_map_blocks_es_recheck(handle_t *handle,
403                                        struct inode *inode,
404                                        struct ext4_map_blocks *es_map,
405                                        struct ext4_map_blocks *map,
406                                        int flags)
407 {
408         int retval;
409
410         map->m_flags = 0;
411         /*
412          * There is a race window that the result is not the same.
413          * e.g. xfstests #223 when dioread_nolock enables.  The reason
414          * is that we lookup a block mapping in extent status tree with
415          * out taking i_data_sem.  So at the time the unwritten extent
416          * could be converted.
417          */
418         down_read(&EXT4_I(inode)->i_data_sem);
419         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
420                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
421                                              EXT4_GET_BLOCKS_KEEP_SIZE);
422         } else {
423                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
424                                              EXT4_GET_BLOCKS_KEEP_SIZE);
425         }
426         up_read((&EXT4_I(inode)->i_data_sem));
427
428         /*
429          * We don't check m_len because extent will be collpased in status
430          * tree.  So the m_len might not equal.
431          */
432         if (es_map->m_lblk != map->m_lblk ||
433             es_map->m_flags != map->m_flags ||
434             es_map->m_pblk != map->m_pblk) {
435                 printk("ES cache assertion failed for inode: %lu "
436                        "es_cached ex [%d/%d/%llu/%x] != "
437                        "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
438                        inode->i_ino, es_map->m_lblk, es_map->m_len,
439                        es_map->m_pblk, es_map->m_flags, map->m_lblk,
440                        map->m_len, map->m_pblk, map->m_flags,
441                        retval, flags);
442         }
443 }
444 #endif /* ES_AGGRESSIVE_TEST */
445
446 /*
447  * The ext4_map_blocks() function tries to look up the requested blocks,
448  * and returns if the blocks are already mapped.
449  *
450  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
451  * and store the allocated blocks in the result buffer head and mark it
452  * mapped.
453  *
454  * If file type is extents based, it will call ext4_ext_map_blocks(),
455  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
456  * based files
457  *
458  * On success, it returns the number of blocks being mapped or allocated.  if
459  * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
460  * is marked as unwritten. If the create == 1, it will mark @map as mapped.
461  *
462  * It returns 0 if plain look up failed (blocks have not been allocated), in
463  * that case, @map is returned as unmapped but we still do fill map->m_len to
464  * indicate the length of a hole starting at map->m_lblk.
465  *
466  * It returns the error in case of allocation failure.
467  */
468 int ext4_map_blocks(handle_t *handle, struct inode *inode,
469                     struct ext4_map_blocks *map, int flags)
470 {
471         struct extent_status es;
472         int retval;
473         int ret = 0;
474 #ifdef ES_AGGRESSIVE_TEST
475         struct ext4_map_blocks orig_map;
476
477         memcpy(&orig_map, map, sizeof(*map));
478 #endif
479
480         map->m_flags = 0;
481         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
482                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
483                   (unsigned long) map->m_lblk);
484
485         /*
486          * ext4_map_blocks returns an int, and m_len is an unsigned int
487          */
488         if (unlikely(map->m_len > INT_MAX))
489                 map->m_len = INT_MAX;
490
491         /* We can handle the block number less than EXT_MAX_BLOCKS */
492         if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
493                 return -EFSCORRUPTED;
494
495         /* Lookup extent status tree firstly */
496         if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
497                 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
498                         map->m_pblk = ext4_es_pblock(&es) +
499                                         map->m_lblk - es.es_lblk;
500                         map->m_flags |= ext4_es_is_written(&es) ?
501                                         EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
502                         retval = es.es_len - (map->m_lblk - es.es_lblk);
503                         if (retval > map->m_len)
504                                 retval = map->m_len;
505                         map->m_len = retval;
506                 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
507                         map->m_pblk = 0;
508                         retval = es.es_len - (map->m_lblk - es.es_lblk);
509                         if (retval > map->m_len)
510                                 retval = map->m_len;
511                         map->m_len = retval;
512                         retval = 0;
513                 } else {
514                         BUG_ON(1);
515                 }
516 #ifdef ES_AGGRESSIVE_TEST
517                 ext4_map_blocks_es_recheck(handle, inode, map,
518                                            &orig_map, flags);
519 #endif
520                 goto found;
521         }
522
523         /*
524          * Try to see if we can get the block without requesting a new
525          * file system block.
526          */
527         down_read(&EXT4_I(inode)->i_data_sem);
528         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
529                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
530                                              EXT4_GET_BLOCKS_KEEP_SIZE);
531         } else {
532                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
533                                              EXT4_GET_BLOCKS_KEEP_SIZE);
534         }
535         if (retval > 0) {
536                 unsigned int status;
537
538                 if (unlikely(retval != map->m_len)) {
539                         ext4_warning(inode->i_sb,
540                                      "ES len assertion failed for inode "
541                                      "%lu: retval %d != map->m_len %d",
542                                      inode->i_ino, retval, map->m_len);
543                         WARN_ON(1);
544                 }
545
546                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
547                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
548                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
549                     !(status & EXTENT_STATUS_WRITTEN) &&
550                     ext4_find_delalloc_range(inode, map->m_lblk,
551                                              map->m_lblk + map->m_len - 1))
552                         status |= EXTENT_STATUS_DELAYED;
553                 ret = ext4_es_insert_extent(inode, map->m_lblk,
554                                             map->m_len, map->m_pblk, status);
555                 if (ret < 0)
556                         retval = ret;
557         }
558         up_read((&EXT4_I(inode)->i_data_sem));
559
560 found:
561         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
562                 ret = check_block_validity(inode, map);
563                 if (ret != 0)
564                         return ret;
565         }
566
567         /* If it is only a block(s) look up */
568         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
569                 return retval;
570
571         /*
572          * Returns if the blocks have already allocated
573          *
574          * Note that if blocks have been preallocated
575          * ext4_ext_get_block() returns the create = 0
576          * with buffer head unmapped.
577          */
578         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
579                 /*
580                  * If we need to convert extent to unwritten
581                  * we continue and do the actual work in
582                  * ext4_ext_map_blocks()
583                  */
584                 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
585                         return retval;
586
587         /*
588          * Here we clear m_flags because after allocating an new extent,
589          * it will be set again.
590          */
591         map->m_flags &= ~EXT4_MAP_FLAGS;
592
593         /*
594          * New blocks allocate and/or writing to unwritten extent
595          * will possibly result in updating i_data, so we take
596          * the write lock of i_data_sem, and call get_block()
597          * with create == 1 flag.
598          */
599         down_write(&EXT4_I(inode)->i_data_sem);
600
601         /*
602          * We need to check for EXT4 here because migrate
603          * could have changed the inode type in between
604          */
605         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
606                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
607         } else {
608                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
609
610                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
611                         /*
612                          * We allocated new blocks which will result in
613                          * i_data's format changing.  Force the migrate
614                          * to fail by clearing migrate flags
615                          */
616                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
617                 }
618
619                 /*
620                  * Update reserved blocks/metadata blocks after successful
621                  * block allocation which had been deferred till now. We don't
622                  * support fallocate for non extent files. So we can update
623                  * reserve space here.
624                  */
625                 if ((retval > 0) &&
626                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
627                         ext4_da_update_reserve_space(inode, retval, 1);
628         }
629
630         if (retval > 0) {
631                 unsigned int status;
632
633                 if (unlikely(retval != map->m_len)) {
634                         ext4_warning(inode->i_sb,
635                                      "ES len assertion failed for inode "
636                                      "%lu: retval %d != map->m_len %d",
637                                      inode->i_ino, retval, map->m_len);
638                         WARN_ON(1);
639                 }
640
641                 /*
642                  * We have to zeroout blocks before inserting them into extent
643                  * status tree. Otherwise someone could look them up there and
644                  * use them before they are really zeroed.
645                  */
646                 if (flags & EXT4_GET_BLOCKS_ZERO &&
647                     map->m_flags & EXT4_MAP_MAPPED &&
648                     map->m_flags & EXT4_MAP_NEW) {
649                         ret = ext4_issue_zeroout(inode, map->m_lblk,
650                                                  map->m_pblk, map->m_len);
651                         if (ret) {
652                                 retval = ret;
653                                 goto out_sem;
654                         }
655                 }
656
657                 /*
658                  * If the extent has been zeroed out, we don't need to update
659                  * extent status tree.
660                  */
661                 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
662                     ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
663                         if (ext4_es_is_written(&es))
664                                 goto out_sem;
665                 }
666                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
667                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
668                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
669                     !(status & EXTENT_STATUS_WRITTEN) &&
670                     ext4_find_delalloc_range(inode, map->m_lblk,
671                                              map->m_lblk + map->m_len - 1))
672                         status |= EXTENT_STATUS_DELAYED;
673                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
674                                             map->m_pblk, status);
675                 if (ret < 0) {
676                         retval = ret;
677                         goto out_sem;
678                 }
679         }
680
681 out_sem:
682         up_write((&EXT4_I(inode)->i_data_sem));
683         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
684                 ret = check_block_validity(inode, map);
685                 if (ret != 0)
686                         return ret;
687
688                 /*
689                  * Inodes with freshly allocated blocks where contents will be
690                  * visible after transaction commit must be on transaction's
691                  * ordered data list.
692                  */
693                 if (map->m_flags & EXT4_MAP_NEW &&
694                     !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
695                     !(flags & EXT4_GET_BLOCKS_ZERO) &&
696                     !IS_NOQUOTA(inode) &&
697                     ext4_should_order_data(inode)) {
698                         ret = ext4_jbd2_file_inode(handle, inode);
699                         if (ret)
700                                 return ret;
701                 }
702         }
703         return retval;
704 }
705
706 /*
707  * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
708  * we have to be careful as someone else may be manipulating b_state as well.
709  */
710 static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
711 {
712         unsigned long old_state;
713         unsigned long new_state;
714
715         flags &= EXT4_MAP_FLAGS;
716
717         /* Dummy buffer_head? Set non-atomically. */
718         if (!bh->b_page) {
719                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
720                 return;
721         }
722         /*
723          * Someone else may be modifying b_state. Be careful! This is ugly but
724          * once we get rid of using bh as a container for mapping information
725          * to pass to / from get_block functions, this can go away.
726          */
727         do {
728                 old_state = READ_ONCE(bh->b_state);
729                 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
730         } while (unlikely(
731                  cmpxchg(&bh->b_state, old_state, new_state) != old_state));
732 }
733
734 static int _ext4_get_block(struct inode *inode, sector_t iblock,
735                            struct buffer_head *bh, int flags)
736 {
737         struct ext4_map_blocks map;
738         int ret = 0;
739
740         if (ext4_has_inline_data(inode))
741                 return -ERANGE;
742
743         map.m_lblk = iblock;
744         map.m_len = bh->b_size >> inode->i_blkbits;
745
746         ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
747                               flags);
748         if (ret > 0) {
749                 map_bh(bh, inode->i_sb, map.m_pblk);
750                 ext4_update_bh_state(bh, map.m_flags);
751                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
752                 ret = 0;
753         }
754         return ret;
755 }
756
757 int ext4_get_block(struct inode *inode, sector_t iblock,
758                    struct buffer_head *bh, int create)
759 {
760         return _ext4_get_block(inode, iblock, bh,
761                                create ? EXT4_GET_BLOCKS_CREATE : 0);
762 }
763
764 /*
765  * Get block function used when preparing for buffered write if we require
766  * creating an unwritten extent if blocks haven't been allocated.  The extent
767  * will be converted to written after the IO is complete.
768  */
769 int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
770                              struct buffer_head *bh_result, int create)
771 {
772         ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
773                    inode->i_ino, create);
774         return _ext4_get_block(inode, iblock, bh_result,
775                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
776 }
777
778 /* Maximum number of blocks we map for direct IO at once. */
779 #define DIO_MAX_BLOCKS 4096
780
781 /*
782  * Get blocks function for the cases that need to start a transaction -
783  * generally difference cases of direct IO and DAX IO. It also handles retries
784  * in case of ENOSPC.
785  */
786 static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
787                                 struct buffer_head *bh_result, int flags)
788 {
789         int dio_credits;
790         handle_t *handle;
791         int retries = 0;
792         int ret;
793
794         /* Trim mapping request to maximum we can map at once for DIO */
795         if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
796                 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
797         dio_credits = ext4_chunk_trans_blocks(inode,
798                                       bh_result->b_size >> inode->i_blkbits);
799 retry:
800         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
801         if (IS_ERR(handle))
802                 return PTR_ERR(handle);
803
804         ret = _ext4_get_block(inode, iblock, bh_result, flags);
805         ext4_journal_stop(handle);
806
807         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
808                 goto retry;
809         return ret;
810 }
811
812 /* Get block function for DIO reads and writes to inodes without extents */
813 int ext4_dio_get_block(struct inode *inode, sector_t iblock,
814                        struct buffer_head *bh, int create)
815 {
816         /* We don't expect handle for direct IO */
817         WARN_ON_ONCE(ext4_journal_current_handle());
818
819         if (!create)
820                 return _ext4_get_block(inode, iblock, bh, 0);
821         return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
822 }
823
824 /*
825  * Get block function for AIO DIO writes when we create unwritten extent if
826  * blocks are not allocated yet. The extent will be converted to written
827  * after IO is complete.
828  */
829 static int ext4_dio_get_block_unwritten_async(struct inode *inode,
830                 sector_t iblock, struct buffer_head *bh_result, int create)
831 {
832         int ret;
833
834         /* We don't expect handle for direct IO */
835         WARN_ON_ONCE(ext4_journal_current_handle());
836
837         ret = ext4_get_block_trans(inode, iblock, bh_result,
838                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
839
840         /*
841          * When doing DIO using unwritten extents, we need io_end to convert
842          * unwritten extents to written on IO completion. We allocate io_end
843          * once we spot unwritten extent and store it in b_private. Generic
844          * DIO code keeps b_private set and furthermore passes the value to
845          * our completion callback in 'private' argument.
846          */
847         if (!ret && buffer_unwritten(bh_result)) {
848                 if (!bh_result->b_private) {
849                         ext4_io_end_t *io_end;
850
851                         io_end = ext4_init_io_end(inode, GFP_KERNEL);
852                         if (!io_end)
853                                 return -ENOMEM;
854                         bh_result->b_private = io_end;
855                         ext4_set_io_unwritten_flag(inode, io_end);
856                 }
857                 set_buffer_defer_completion(bh_result);
858         }
859
860         return ret;
861 }
862
863 /*
864  * Get block function for non-AIO DIO writes when we create unwritten extent if
865  * blocks are not allocated yet. The extent will be converted to written
866  * after IO is complete from ext4_ext_direct_IO() function.
867  */
868 static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
869                 sector_t iblock, struct buffer_head *bh_result, int create)
870 {
871         int ret;
872
873         /* We don't expect handle for direct IO */
874         WARN_ON_ONCE(ext4_journal_current_handle());
875
876         ret = ext4_get_block_trans(inode, iblock, bh_result,
877                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
878
879         /*
880          * Mark inode as having pending DIO writes to unwritten extents.
881          * ext4_ext_direct_IO() checks this flag and converts extents to
882          * written.
883          */
884         if (!ret && buffer_unwritten(bh_result))
885                 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
886
887         return ret;
888 }
889
890 static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
891                    struct buffer_head *bh_result, int create)
892 {
893         int ret;
894
895         ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
896                    inode->i_ino, create);
897         /* We don't expect handle for direct IO */
898         WARN_ON_ONCE(ext4_journal_current_handle());
899
900         ret = _ext4_get_block(inode, iblock, bh_result, 0);
901         /*
902          * Blocks should have been preallocated! ext4_file_write_iter() checks
903          * that.
904          */
905         WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
906
907         return ret;
908 }
909
910
911 /*
912  * `handle' can be NULL if create is zero
913  */
914 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
915                                 ext4_lblk_t block, int map_flags)
916 {
917         struct ext4_map_blocks map;
918         struct buffer_head *bh;
919         int create = map_flags & EXT4_GET_BLOCKS_CREATE;
920         int err;
921
922         J_ASSERT(handle != NULL || create == 0);
923
924         map.m_lblk = block;
925         map.m_len = 1;
926         err = ext4_map_blocks(handle, inode, &map, map_flags);
927
928         if (err == 0)
929                 return create ? ERR_PTR(-ENOSPC) : NULL;
930         if (err < 0)
931                 return ERR_PTR(err);
932
933         bh = sb_getblk(inode->i_sb, map.m_pblk);
934         if (unlikely(!bh))
935                 return ERR_PTR(-ENOMEM);
936         if (map.m_flags & EXT4_MAP_NEW) {
937                 J_ASSERT(create != 0);
938                 J_ASSERT(handle != NULL);
939
940                 /*
941                  * Now that we do not always journal data, we should
942                  * keep in mind whether this should always journal the
943                  * new buffer as metadata.  For now, regular file
944                  * writes use ext4_get_block instead, so it's not a
945                  * problem.
946                  */
947                 lock_buffer(bh);
948                 BUFFER_TRACE(bh, "call get_create_access");
949                 err = ext4_journal_get_create_access(handle, bh);
950                 if (unlikely(err)) {
951                         unlock_buffer(bh);
952                         goto errout;
953                 }
954                 if (!buffer_uptodate(bh)) {
955                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
956                         set_buffer_uptodate(bh);
957                 }
958                 unlock_buffer(bh);
959                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
960                 err = ext4_handle_dirty_metadata(handle, inode, bh);
961                 if (unlikely(err))
962                         goto errout;
963         } else
964                 BUFFER_TRACE(bh, "not a new buffer");
965         return bh;
966 errout:
967         brelse(bh);
968         return ERR_PTR(err);
969 }
970
971 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
972                                ext4_lblk_t block, int map_flags)
973 {
974         struct buffer_head *bh;
975
976         bh = ext4_getblk(handle, inode, block, map_flags);
977         if (IS_ERR(bh))
978                 return bh;
979         if (!bh || buffer_uptodate(bh))
980                 return bh;
981         ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
982         wait_on_buffer(bh);
983         if (buffer_uptodate(bh))
984                 return bh;
985         put_bh(bh);
986         return ERR_PTR(-EIO);
987 }
988
989 int ext4_walk_page_buffers(handle_t *handle,
990                            struct buffer_head *head,
991                            unsigned from,
992                            unsigned to,
993                            int *partial,
994                            int (*fn)(handle_t *handle,
995                                      struct buffer_head *bh))
996 {
997         struct buffer_head *bh;
998         unsigned block_start, block_end;
999         unsigned blocksize = head->b_size;
1000         int err, ret = 0;
1001         struct buffer_head *next;
1002
1003         for (bh = head, block_start = 0;
1004              ret == 0 && (bh != head || !block_start);
1005              block_start = block_end, bh = next) {
1006                 next = bh->b_this_page;
1007                 block_end = block_start + blocksize;
1008                 if (block_end <= from || block_start >= to) {
1009                         if (partial && !buffer_uptodate(bh))
1010                                 *partial = 1;
1011                         continue;
1012                 }
1013                 err = (*fn)(handle, bh);
1014                 if (!ret)
1015                         ret = err;
1016         }
1017         return ret;
1018 }
1019
1020 /*
1021  * To preserve ordering, it is essential that the hole instantiation and
1022  * the data write be encapsulated in a single transaction.  We cannot
1023  * close off a transaction and start a new one between the ext4_get_block()
1024  * and the commit_write().  So doing the jbd2_journal_start at the start of
1025  * prepare_write() is the right place.
1026  *
1027  * Also, this function can nest inside ext4_writepage().  In that case, we
1028  * *know* that ext4_writepage() has generated enough buffer credits to do the
1029  * whole page.  So we won't block on the journal in that case, which is good,
1030  * because the caller may be PF_MEMALLOC.
1031  *
1032  * By accident, ext4 can be reentered when a transaction is open via
1033  * quota file writes.  If we were to commit the transaction while thus
1034  * reentered, there can be a deadlock - we would be holding a quota
1035  * lock, and the commit would never complete if another thread had a
1036  * transaction open and was blocking on the quota lock - a ranking
1037  * violation.
1038  *
1039  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1040  * will _not_ run commit under these circumstances because handle->h_ref
1041  * is elevated.  We'll still have enough credits for the tiny quotafile
1042  * write.
1043  */
1044 int do_journal_get_write_access(handle_t *handle,
1045                                 struct buffer_head *bh)
1046 {
1047         int dirty = buffer_dirty(bh);
1048         int ret;
1049
1050         if (!buffer_mapped(bh) || buffer_freed(bh))
1051                 return 0;
1052         /*
1053          * __block_write_begin() could have dirtied some buffers. Clean
1054          * the dirty bit as jbd2_journal_get_write_access() could complain
1055          * otherwise about fs integrity issues. Setting of the dirty bit
1056          * by __block_write_begin() isn't a real problem here as we clear
1057          * the bit before releasing a page lock and thus writeback cannot
1058          * ever write the buffer.
1059          */
1060         if (dirty)
1061                 clear_buffer_dirty(bh);
1062         BUFFER_TRACE(bh, "get write access");
1063         ret = ext4_journal_get_write_access(handle, bh);
1064         if (!ret && dirty)
1065                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1066         return ret;
1067 }
1068
1069 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1070 static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1071                                   get_block_t *get_block)
1072 {
1073         unsigned from = pos & (PAGE_SIZE - 1);
1074         unsigned to = from + len;
1075         struct inode *inode = page->mapping->host;
1076         unsigned block_start, block_end;
1077         sector_t block;
1078         int err = 0;
1079         unsigned blocksize = inode->i_sb->s_blocksize;
1080         unsigned bbits;
1081         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1082         bool decrypt = false;
1083
1084         BUG_ON(!PageLocked(page));
1085         BUG_ON(from > PAGE_SIZE);
1086         BUG_ON(to > PAGE_SIZE);
1087         BUG_ON(from > to);
1088
1089         if (!page_has_buffers(page))
1090                 create_empty_buffers(page, blocksize, 0);
1091         head = page_buffers(page);
1092         bbits = ilog2(blocksize);
1093         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1094
1095         for (bh = head, block_start = 0; bh != head || !block_start;
1096             block++, block_start = block_end, bh = bh->b_this_page) {
1097                 block_end = block_start + blocksize;
1098                 if (block_end <= from || block_start >= to) {
1099                         if (PageUptodate(page)) {
1100                                 if (!buffer_uptodate(bh))
1101                                         set_buffer_uptodate(bh);
1102                         }
1103                         continue;
1104                 }
1105                 if (buffer_new(bh))
1106                         clear_buffer_new(bh);
1107                 if (!buffer_mapped(bh)) {
1108                         WARN_ON(bh->b_size != blocksize);
1109                         err = get_block(inode, block, bh, 1);
1110                         if (err)
1111                                 break;
1112                         if (buffer_new(bh)) {
1113                                 unmap_underlying_metadata(bh->b_bdev,
1114                                                           bh->b_blocknr);
1115                                 if (PageUptodate(page)) {
1116                                         clear_buffer_new(bh);
1117                                         set_buffer_uptodate(bh);
1118                                         mark_buffer_dirty(bh);
1119                                         continue;
1120                                 }
1121                                 if (block_end > to || block_start < from)
1122                                         zero_user_segments(page, to, block_end,
1123                                                            block_start, from);
1124                                 continue;
1125                         }
1126                 }
1127                 if (PageUptodate(page)) {
1128                         if (!buffer_uptodate(bh))
1129                                 set_buffer_uptodate(bh);
1130                         continue;
1131                 }
1132                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1133                     !buffer_unwritten(bh) &&
1134                     (block_start < from || block_end > to)) {
1135                         ll_rw_block(READ, 1, &bh);
1136                         *wait_bh++ = bh;
1137                         decrypt = ext4_encrypted_inode(inode) &&
1138                                 S_ISREG(inode->i_mode);
1139                 }
1140         }
1141         /*
1142          * If we issued read requests, let them complete.
1143          */
1144         while (wait_bh > wait) {
1145                 wait_on_buffer(*--wait_bh);
1146                 if (!buffer_uptodate(*wait_bh))
1147                         err = -EIO;
1148         }
1149         if (unlikely(err))
1150                 page_zero_new_buffers(page, from, to);
1151         else if (decrypt)
1152                 err = ext4_decrypt(page);
1153         return err;
1154 }
1155 #endif
1156
1157 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1158                             loff_t pos, unsigned len, unsigned flags,
1159                             struct page **pagep, void **fsdata)
1160 {
1161         struct inode *inode = mapping->host;
1162         int ret, needed_blocks;
1163         handle_t *handle;
1164         int retries = 0;
1165         struct page *page;
1166         pgoff_t index;
1167         unsigned from, to;
1168
1169         trace_ext4_write_begin(inode, pos, len, flags);
1170         /*
1171          * Reserve one block more for addition to orphan list in case
1172          * we allocate blocks but write fails for some reason
1173          */
1174         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1175         index = pos >> PAGE_SHIFT;
1176         from = pos & (PAGE_SIZE - 1);
1177         to = from + len;
1178
1179         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1180                 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1181                                                     flags, pagep);
1182                 if (ret < 0)
1183                         return ret;
1184                 if (ret == 1)
1185                         return 0;
1186         }
1187
1188         /*
1189          * grab_cache_page_write_begin() can take a long time if the
1190          * system is thrashing due to memory pressure, or if the page
1191          * is being written back.  So grab it first before we start
1192          * the transaction handle.  This also allows us to allocate
1193          * the page (if needed) without using GFP_NOFS.
1194          */
1195 retry_grab:
1196         page = grab_cache_page_write_begin(mapping, index, flags);
1197         if (!page)
1198                 return -ENOMEM;
1199         unlock_page(page);
1200
1201 retry_journal:
1202         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1203         if (IS_ERR(handle)) {
1204                 put_page(page);
1205                 return PTR_ERR(handle);
1206         }
1207
1208         lock_page(page);
1209         if (page->mapping != mapping) {
1210                 /* The page got truncated from under us */
1211                 unlock_page(page);
1212                 put_page(page);
1213                 ext4_journal_stop(handle);
1214                 goto retry_grab;
1215         }
1216         /* In case writeback began while the page was unlocked */
1217         wait_for_stable_page(page);
1218
1219 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1220         if (ext4_should_dioread_nolock(inode))
1221                 ret = ext4_block_write_begin(page, pos, len,
1222                                              ext4_get_block_unwritten);
1223         else
1224                 ret = ext4_block_write_begin(page, pos, len,
1225                                              ext4_get_block);
1226 #else
1227         if (ext4_should_dioread_nolock(inode))
1228                 ret = __block_write_begin(page, pos, len,
1229                                           ext4_get_block_unwritten);
1230         else
1231                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1232 #endif
1233         if (!ret && ext4_should_journal_data(inode)) {
1234                 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1235                                              from, to, NULL,
1236                                              do_journal_get_write_access);
1237         }
1238
1239         if (ret) {
1240                 unlock_page(page);
1241                 /*
1242                  * __block_write_begin may have instantiated a few blocks
1243                  * outside i_size.  Trim these off again. Don't need
1244                  * i_size_read because we hold i_mutex.
1245                  *
1246                  * Add inode to orphan list in case we crash before
1247                  * truncate finishes
1248                  */
1249                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1250                         ext4_orphan_add(handle, inode);
1251
1252                 ext4_journal_stop(handle);
1253                 if (pos + len > inode->i_size) {
1254                         ext4_truncate_failed_write(inode);
1255                         /*
1256                          * If truncate failed early the inode might
1257                          * still be on the orphan list; we need to
1258                          * make sure the inode is removed from the
1259                          * orphan list in that case.
1260                          */
1261                         if (inode->i_nlink)
1262                                 ext4_orphan_del(NULL, inode);
1263                 }
1264
1265                 if (ret == -ENOSPC &&
1266                     ext4_should_retry_alloc(inode->i_sb, &retries))
1267                         goto retry_journal;
1268                 put_page(page);
1269                 return ret;
1270         }
1271         *pagep = page;
1272         return ret;
1273 }
1274
1275 /* For write_end() in data=journal mode */
1276 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1277 {
1278         int ret;
1279         if (!buffer_mapped(bh) || buffer_freed(bh))
1280                 return 0;
1281         set_buffer_uptodate(bh);
1282         ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1283         clear_buffer_meta(bh);
1284         clear_buffer_prio(bh);
1285         return ret;
1286 }
1287
1288 /*
1289  * We need to pick up the new inode size which generic_commit_write gave us
1290  * `file' can be NULL - eg, when called from page_symlink().
1291  *
1292  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1293  * buffers are managed internally.
1294  */
1295 static int ext4_write_end(struct file *file,
1296                           struct address_space *mapping,
1297                           loff_t pos, unsigned len, unsigned copied,
1298                           struct page *page, void *fsdata)
1299 {
1300         handle_t *handle = ext4_journal_current_handle();
1301         struct inode *inode = mapping->host;
1302         loff_t old_size = inode->i_size;
1303         int ret = 0, ret2;
1304         int i_size_changed = 0;
1305
1306         trace_ext4_write_end(inode, pos, len, copied);
1307         if (ext4_has_inline_data(inode)) {
1308                 ret = ext4_write_inline_data_end(inode, pos, len,
1309                                                  copied, page);
1310                 if (ret < 0)
1311                         goto errout;
1312                 copied = ret;
1313         } else
1314                 copied = block_write_end(file, mapping, pos,
1315                                          len, copied, page, fsdata);
1316         /*
1317          * it's important to update i_size while still holding page lock:
1318          * page writeout could otherwise come in and zero beyond i_size.
1319          */
1320         i_size_changed = ext4_update_inode_size(inode, pos + copied);
1321         unlock_page(page);
1322         put_page(page);
1323
1324         if (old_size < pos)
1325                 pagecache_isize_extended(inode, old_size, pos);
1326         /*
1327          * Don't mark the inode dirty under page lock. First, it unnecessarily
1328          * makes the holding time of page lock longer. Second, it forces lock
1329          * ordering of page lock and transaction start for journaling
1330          * filesystems.
1331          */
1332         if (i_size_changed)
1333                 ext4_mark_inode_dirty(handle, inode);
1334
1335         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1336                 /* if we have allocated more blocks and copied
1337                  * less. We will have blocks allocated outside
1338                  * inode->i_size. So truncate them
1339                  */
1340                 ext4_orphan_add(handle, inode);
1341 errout:
1342         ret2 = ext4_journal_stop(handle);
1343         if (!ret)
1344                 ret = ret2;
1345
1346         if (pos + len > inode->i_size) {
1347                 ext4_truncate_failed_write(inode);
1348                 /*
1349                  * If truncate failed early the inode might still be
1350                  * on the orphan list; we need to make sure the inode
1351                  * is removed from the orphan list in that case.
1352                  */
1353                 if (inode->i_nlink)
1354                         ext4_orphan_del(NULL, inode);
1355         }
1356
1357         return ret ? ret : copied;
1358 }
1359
1360 /*
1361  * This is a private version of page_zero_new_buffers() which doesn't
1362  * set the buffer to be dirty, since in data=journalled mode we need
1363  * to call ext4_handle_dirty_metadata() instead.
1364  */
1365 static void zero_new_buffers(struct page *page, unsigned from, unsigned to)
1366 {
1367         unsigned int block_start = 0, block_end;
1368         struct buffer_head *head, *bh;
1369
1370         bh = head = page_buffers(page);
1371         do {
1372                 block_end = block_start + bh->b_size;
1373                 if (buffer_new(bh)) {
1374                         if (block_end > from && block_start < to) {
1375                                 if (!PageUptodate(page)) {
1376                                         unsigned start, size;
1377
1378                                         start = max(from, block_start);
1379                                         size = min(to, block_end) - start;
1380
1381                                         zero_user(page, start, size);
1382                                         set_buffer_uptodate(bh);
1383                                 }
1384                                 clear_buffer_new(bh);
1385                         }
1386                 }
1387                 block_start = block_end;
1388                 bh = bh->b_this_page;
1389         } while (bh != head);
1390 }
1391
1392 static int ext4_journalled_write_end(struct file *file,
1393                                      struct address_space *mapping,
1394                                      loff_t pos, unsigned len, unsigned copied,
1395                                      struct page *page, void *fsdata)
1396 {
1397         handle_t *handle = ext4_journal_current_handle();
1398         struct inode *inode = mapping->host;
1399         loff_t old_size = inode->i_size;
1400         int ret = 0, ret2;
1401         int partial = 0;
1402         unsigned from, to;
1403         int size_changed = 0;
1404
1405         trace_ext4_journalled_write_end(inode, pos, len, copied);
1406         from = pos & (PAGE_SIZE - 1);
1407         to = from + len;
1408
1409         BUG_ON(!ext4_handle_valid(handle));
1410
1411         if (ext4_has_inline_data(inode))
1412                 copied = ext4_write_inline_data_end(inode, pos, len,
1413                                                     copied, page);
1414         else {
1415                 if (copied < len) {
1416                         if (!PageUptodate(page))
1417                                 copied = 0;
1418                         zero_new_buffers(page, from+copied, to);
1419                 }
1420
1421                 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1422                                              to, &partial, write_end_fn);
1423                 if (!partial)
1424                         SetPageUptodate(page);
1425         }
1426         size_changed = ext4_update_inode_size(inode, pos + copied);
1427         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1428         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1429         unlock_page(page);
1430         put_page(page);
1431
1432         if (old_size < pos)
1433                 pagecache_isize_extended(inode, old_size, pos);
1434
1435         if (size_changed) {
1436                 ret2 = ext4_mark_inode_dirty(handle, inode);
1437                 if (!ret)
1438                         ret = ret2;
1439         }
1440
1441         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1442                 /* if we have allocated more blocks and copied
1443                  * less. We will have blocks allocated outside
1444                  * inode->i_size. So truncate them
1445                  */
1446                 ext4_orphan_add(handle, inode);
1447
1448         ret2 = ext4_journal_stop(handle);
1449         if (!ret)
1450                 ret = ret2;
1451         if (pos + len > inode->i_size) {
1452                 ext4_truncate_failed_write(inode);
1453                 /*
1454                  * If truncate failed early the inode might still be
1455                  * on the orphan list; we need to make sure the inode
1456                  * is removed from the orphan list in that case.
1457                  */
1458                 if (inode->i_nlink)
1459                         ext4_orphan_del(NULL, inode);
1460         }
1461
1462         return ret ? ret : copied;
1463 }
1464
1465 /*
1466  * Reserve space for a single cluster
1467  */
1468 static int ext4_da_reserve_space(struct inode *inode)
1469 {
1470         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1471         struct ext4_inode_info *ei = EXT4_I(inode);
1472         int ret;
1473
1474         /*
1475          * We will charge metadata quota at writeout time; this saves
1476          * us from metadata over-estimation, though we may go over by
1477          * a small amount in the end.  Here we just reserve for data.
1478          */
1479         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1480         if (ret)
1481                 return ret;
1482
1483         spin_lock(&ei->i_block_reservation_lock);
1484         if (ext4_claim_free_clusters(sbi, 1, 0)) {
1485                 spin_unlock(&ei->i_block_reservation_lock);
1486                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1487                 return -ENOSPC;
1488         }
1489         ei->i_reserved_data_blocks++;
1490         trace_ext4_da_reserve_space(inode);
1491         spin_unlock(&ei->i_block_reservation_lock);
1492
1493         return 0;       /* success */
1494 }
1495
1496 static void ext4_da_release_space(struct inode *inode, int to_free)
1497 {
1498         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1499         struct ext4_inode_info *ei = EXT4_I(inode);
1500
1501         if (!to_free)
1502                 return;         /* Nothing to release, exit */
1503
1504         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1505
1506         trace_ext4_da_release_space(inode, to_free);
1507         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1508                 /*
1509                  * if there aren't enough reserved blocks, then the
1510                  * counter is messed up somewhere.  Since this
1511                  * function is called from invalidate page, it's
1512                  * harmless to return without any action.
1513                  */
1514                 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1515                          "ino %lu, to_free %d with only %d reserved "
1516                          "data blocks", inode->i_ino, to_free,
1517                          ei->i_reserved_data_blocks);
1518                 WARN_ON(1);
1519                 to_free = ei->i_reserved_data_blocks;
1520         }
1521         ei->i_reserved_data_blocks -= to_free;
1522
1523         /* update fs dirty data blocks counter */
1524         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1525
1526         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1527
1528         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1529 }
1530
1531 static void ext4_da_page_release_reservation(struct page *page,
1532                                              unsigned int offset,
1533                                              unsigned int length)
1534 {
1535         int to_release = 0, contiguous_blks = 0;
1536         struct buffer_head *head, *bh;
1537         unsigned int curr_off = 0;
1538         struct inode *inode = page->mapping->host;
1539         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1540         unsigned int stop = offset + length;
1541         int num_clusters;
1542         ext4_fsblk_t lblk;
1543
1544         BUG_ON(stop > PAGE_SIZE || stop < length);
1545
1546         head = page_buffers(page);
1547         bh = head;
1548         do {
1549                 unsigned int next_off = curr_off + bh->b_size;
1550
1551                 if (next_off > stop)
1552                         break;
1553
1554                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1555                         to_release++;
1556                         contiguous_blks++;
1557                         clear_buffer_delay(bh);
1558                 } else if (contiguous_blks) {
1559                         lblk = page->index <<
1560                                (PAGE_SHIFT - inode->i_blkbits);
1561                         lblk += (curr_off >> inode->i_blkbits) -
1562                                 contiguous_blks;
1563                         ext4_es_remove_extent(inode, lblk, contiguous_blks);
1564                         contiguous_blks = 0;
1565                 }
1566                 curr_off = next_off;
1567         } while ((bh = bh->b_this_page) != head);
1568
1569         if (contiguous_blks) {
1570                 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1571                 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1572                 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1573         }
1574
1575         /* If we have released all the blocks belonging to a cluster, then we
1576          * need to release the reserved space for that cluster. */
1577         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1578         while (num_clusters > 0) {
1579                 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1580                         ((num_clusters - 1) << sbi->s_cluster_bits);
1581                 if (sbi->s_cluster_ratio == 1 ||
1582                     !ext4_find_delalloc_cluster(inode, lblk))
1583                         ext4_da_release_space(inode, 1);
1584
1585                 num_clusters--;
1586         }
1587 }
1588
1589 /*
1590  * Delayed allocation stuff
1591  */
1592
1593 struct mpage_da_data {
1594         struct inode *inode;
1595         struct writeback_control *wbc;
1596
1597         pgoff_t first_page;     /* The first page to write */
1598         pgoff_t next_page;      /* Current page to examine */
1599         pgoff_t last_page;      /* Last page to examine */
1600         /*
1601          * Extent to map - this can be after first_page because that can be
1602          * fully mapped. We somewhat abuse m_flags to store whether the extent
1603          * is delalloc or unwritten.
1604          */
1605         struct ext4_map_blocks map;
1606         struct ext4_io_submit io_submit;        /* IO submission data */
1607 };
1608
1609 static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1610                                        bool invalidate)
1611 {
1612         int nr_pages, i;
1613         pgoff_t index, end;
1614         struct pagevec pvec;
1615         struct inode *inode = mpd->inode;
1616         struct address_space *mapping = inode->i_mapping;
1617
1618         /* This is necessary when next_page == 0. */
1619         if (mpd->first_page >= mpd->next_page)
1620                 return;
1621
1622         index = mpd->first_page;
1623         end   = mpd->next_page - 1;
1624         if (invalidate) {
1625                 ext4_lblk_t start, last;
1626                 start = index << (PAGE_SHIFT - inode->i_blkbits);
1627                 last = end << (PAGE_SHIFT - inode->i_blkbits);
1628                 ext4_es_remove_extent(inode, start, last - start + 1);
1629         }
1630
1631         pagevec_init(&pvec, 0);
1632         while (index <= end) {
1633                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1634                 if (nr_pages == 0)
1635                         break;
1636                 for (i = 0; i < nr_pages; i++) {
1637                         struct page *page = pvec.pages[i];
1638                         if (page->index > end)
1639                                 break;
1640                         BUG_ON(!PageLocked(page));
1641                         BUG_ON(PageWriteback(page));
1642                         if (invalidate) {
1643                                 block_invalidatepage(page, 0, PAGE_SIZE);
1644                                 ClearPageUptodate(page);
1645                         }
1646                         unlock_page(page);
1647                 }
1648                 index = pvec.pages[nr_pages - 1]->index + 1;
1649                 pagevec_release(&pvec);
1650         }
1651 }
1652
1653 static void ext4_print_free_blocks(struct inode *inode)
1654 {
1655         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1656         struct super_block *sb = inode->i_sb;
1657         struct ext4_inode_info *ei = EXT4_I(inode);
1658
1659         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1660                EXT4_C2B(EXT4_SB(inode->i_sb),
1661                         ext4_count_free_clusters(sb)));
1662         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1663         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1664                (long long) EXT4_C2B(EXT4_SB(sb),
1665                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1666         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1667                (long long) EXT4_C2B(EXT4_SB(sb),
1668                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1669         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1670         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1671                  ei->i_reserved_data_blocks);
1672         return;
1673 }
1674
1675 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1676 {
1677         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1678 }
1679
1680 /*
1681  * This function is grabs code from the very beginning of
1682  * ext4_map_blocks, but assumes that the caller is from delayed write
1683  * time. This function looks up the requested blocks and sets the
1684  * buffer delay bit under the protection of i_data_sem.
1685  */
1686 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1687                               struct ext4_map_blocks *map,
1688                               struct buffer_head *bh)
1689 {
1690         struct extent_status es;
1691         int retval;
1692         sector_t invalid_block = ~((sector_t) 0xffff);
1693 #ifdef ES_AGGRESSIVE_TEST
1694         struct ext4_map_blocks orig_map;
1695
1696         memcpy(&orig_map, map, sizeof(*map));
1697 #endif
1698
1699         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1700                 invalid_block = ~0;
1701
1702         map->m_flags = 0;
1703         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1704                   "logical block %lu\n", inode->i_ino, map->m_len,
1705                   (unsigned long) map->m_lblk);
1706
1707         /* Lookup extent status tree firstly */
1708         if (ext4_es_lookup_extent(inode, iblock, &es)) {
1709                 if (ext4_es_is_hole(&es)) {
1710                         retval = 0;
1711                         down_read(&EXT4_I(inode)->i_data_sem);
1712                         goto add_delayed;
1713                 }
1714
1715                 /*
1716                  * Delayed extent could be allocated by fallocate.
1717                  * So we need to check it.
1718                  */
1719                 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1720                         map_bh(bh, inode->i_sb, invalid_block);
1721                         set_buffer_new(bh);
1722                         set_buffer_delay(bh);
1723                         return 0;
1724                 }
1725
1726                 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1727                 retval = es.es_len - (iblock - es.es_lblk);
1728                 if (retval > map->m_len)
1729                         retval = map->m_len;
1730                 map->m_len = retval;
1731                 if (ext4_es_is_written(&es))
1732                         map->m_flags |= EXT4_MAP_MAPPED;
1733                 else if (ext4_es_is_unwritten(&es))
1734                         map->m_flags |= EXT4_MAP_UNWRITTEN;
1735                 else
1736                         BUG_ON(1);
1737
1738 #ifdef ES_AGGRESSIVE_TEST
1739                 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1740 #endif
1741                 return retval;
1742         }
1743
1744         /*
1745          * Try to see if we can get the block without requesting a new
1746          * file system block.
1747          */
1748         down_read(&EXT4_I(inode)->i_data_sem);
1749         if (ext4_has_inline_data(inode))
1750                 retval = 0;
1751         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1752                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1753         else
1754                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1755
1756 add_delayed:
1757         if (retval == 0) {
1758                 int ret;
1759                 /*
1760                  * XXX: __block_prepare_write() unmaps passed block,
1761                  * is it OK?
1762                  */
1763                 /*
1764                  * If the block was allocated from previously allocated cluster,
1765                  * then we don't need to reserve it again. However we still need
1766                  * to reserve metadata for every block we're going to write.
1767                  */
1768                 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1769                     !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1770                         ret = ext4_da_reserve_space(inode);
1771                         if (ret) {
1772                                 /* not enough space to reserve */
1773                                 retval = ret;
1774                                 goto out_unlock;
1775                         }
1776                 }
1777
1778                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1779                                             ~0, EXTENT_STATUS_DELAYED);
1780                 if (ret) {
1781                         retval = ret;
1782                         goto out_unlock;
1783                 }
1784
1785                 map_bh(bh, inode->i_sb, invalid_block);
1786                 set_buffer_new(bh);
1787                 set_buffer_delay(bh);
1788         } else if (retval > 0) {
1789                 int ret;
1790                 unsigned int status;
1791
1792                 if (unlikely(retval != map->m_len)) {
1793                         ext4_warning(inode->i_sb,
1794                                      "ES len assertion failed for inode "
1795                                      "%lu: retval %d != map->m_len %d",
1796                                      inode->i_ino, retval, map->m_len);
1797                         WARN_ON(1);
1798                 }
1799
1800                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1801                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1802                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1803                                             map->m_pblk, status);
1804                 if (ret != 0)
1805                         retval = ret;
1806         }
1807
1808 out_unlock:
1809         up_read((&EXT4_I(inode)->i_data_sem));
1810
1811         return retval;
1812 }
1813
1814 /*
1815  * This is a special get_block_t callback which is used by
1816  * ext4_da_write_begin().  It will either return mapped block or
1817  * reserve space for a single block.
1818  *
1819  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1820  * We also have b_blocknr = -1 and b_bdev initialized properly
1821  *
1822  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1823  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1824  * initialized properly.
1825  */
1826 int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1827                            struct buffer_head *bh, int create)
1828 {
1829         struct ext4_map_blocks map;
1830         int ret = 0;
1831
1832         BUG_ON(create == 0);
1833         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1834
1835         map.m_lblk = iblock;
1836         map.m_len = 1;
1837
1838         /*
1839          * first, we need to know whether the block is allocated already
1840          * preallocated blocks are unmapped but should treated
1841          * the same as allocated blocks.
1842          */
1843         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1844         if (ret <= 0)
1845                 return ret;
1846
1847         map_bh(bh, inode->i_sb, map.m_pblk);
1848         ext4_update_bh_state(bh, map.m_flags);
1849
1850         if (buffer_unwritten(bh)) {
1851                 /* A delayed write to unwritten bh should be marked
1852                  * new and mapped.  Mapped ensures that we don't do
1853                  * get_block multiple times when we write to the same
1854                  * offset and new ensures that we do proper zero out
1855                  * for partial write.
1856                  */
1857                 set_buffer_new(bh);
1858                 set_buffer_mapped(bh);
1859         }
1860         return 0;
1861 }
1862
1863 static int bget_one(handle_t *handle, struct buffer_head *bh)
1864 {
1865         get_bh(bh);
1866         return 0;
1867 }
1868
1869 static int bput_one(handle_t *handle, struct buffer_head *bh)
1870 {
1871         put_bh(bh);
1872         return 0;
1873 }
1874
1875 static int __ext4_journalled_writepage(struct page *page,
1876                                        unsigned int len)
1877 {
1878         struct address_space *mapping = page->mapping;
1879         struct inode *inode = mapping->host;
1880         struct buffer_head *page_bufs = NULL;
1881         handle_t *handle = NULL;
1882         int ret = 0, err = 0;
1883         int inline_data = ext4_has_inline_data(inode);
1884         struct buffer_head *inode_bh = NULL;
1885
1886         ClearPageChecked(page);
1887
1888         if (inline_data) {
1889                 BUG_ON(page->index != 0);
1890                 BUG_ON(len > ext4_get_max_inline_size(inode));
1891                 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1892                 if (inode_bh == NULL)
1893                         goto out;
1894         } else {
1895                 page_bufs = page_buffers(page);
1896                 if (!page_bufs) {
1897                         BUG();
1898                         goto out;
1899                 }
1900                 ext4_walk_page_buffers(handle, page_bufs, 0, len,
1901                                        NULL, bget_one);
1902         }
1903         /*
1904          * We need to release the page lock before we start the
1905          * journal, so grab a reference so the page won't disappear
1906          * out from under us.
1907          */
1908         get_page(page);
1909         unlock_page(page);
1910
1911         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
1912                                     ext4_writepage_trans_blocks(inode));
1913         if (IS_ERR(handle)) {
1914                 ret = PTR_ERR(handle);
1915                 put_page(page);
1916                 goto out_no_pagelock;
1917         }
1918         BUG_ON(!ext4_handle_valid(handle));
1919
1920         lock_page(page);
1921         put_page(page);
1922         if (page->mapping != mapping) {
1923                 /* The page got truncated from under us */
1924                 ext4_journal_stop(handle);
1925                 ret = 0;
1926                 goto out;
1927         }
1928
1929         if (inline_data) {
1930                 BUFFER_TRACE(inode_bh, "get write access");
1931                 ret = ext4_journal_get_write_access(handle, inode_bh);
1932
1933                 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
1934
1935         } else {
1936                 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1937                                              do_journal_get_write_access);
1938
1939                 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1940                                              write_end_fn);
1941         }
1942         if (ret == 0)
1943                 ret = err;
1944         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1945         err = ext4_journal_stop(handle);
1946         if (!ret)
1947                 ret = err;
1948
1949         if (!ext4_has_inline_data(inode))
1950                 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
1951                                        NULL, bput_one);
1952         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1953 out:
1954         unlock_page(page);
1955 out_no_pagelock:
1956         brelse(inode_bh);
1957         return ret;
1958 }
1959
1960 /*
1961  * Note that we don't need to start a transaction unless we're journaling data
1962  * because we should have holes filled from ext4_page_mkwrite(). We even don't
1963  * need to file the inode to the transaction's list in ordered mode because if
1964  * we are writing back data added by write(), the inode is already there and if
1965  * we are writing back data modified via mmap(), no one guarantees in which
1966  * transaction the data will hit the disk. In case we are journaling data, we
1967  * cannot start transaction directly because transaction start ranks above page
1968  * lock so we have to do some magic.
1969  *
1970  * This function can get called via...
1971  *   - ext4_writepages after taking page lock (have journal handle)
1972  *   - journal_submit_inode_data_buffers (no journal handle)
1973  *   - shrink_page_list via the kswapd/direct reclaim (no journal handle)
1974  *   - grab_page_cache when doing write_begin (have journal handle)
1975  *
1976  * We don't do any block allocation in this function. If we have page with
1977  * multiple blocks we need to write those buffer_heads that are mapped. This
1978  * is important for mmaped based write. So if we do with blocksize 1K
1979  * truncate(f, 1024);
1980  * a = mmap(f, 0, 4096);
1981  * a[0] = 'a';
1982  * truncate(f, 4096);
1983  * we have in the page first buffer_head mapped via page_mkwrite call back
1984  * but other buffer_heads would be unmapped but dirty (dirty done via the
1985  * do_wp_page). So writepage should write the first block. If we modify
1986  * the mmap area beyond 1024 we will again get a page_fault and the
1987  * page_mkwrite callback will do the block allocation and mark the
1988  * buffer_heads mapped.
1989  *
1990  * We redirty the page if we have any buffer_heads that is either delay or
1991  * unwritten in the page.
1992  *
1993  * We can get recursively called as show below.
1994  *
1995  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1996  *              ext4_writepage()
1997  *
1998  * But since we don't do any block allocation we should not deadlock.
1999  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2000  */
2001 static int ext4_writepage(struct page *page,
2002                           struct writeback_control *wbc)
2003 {
2004         int ret = 0;
2005         loff_t size;
2006         unsigned int len;
2007         struct buffer_head *page_bufs = NULL;
2008         struct inode *inode = page->mapping->host;
2009         struct ext4_io_submit io_submit;
2010         bool keep_towrite = false;
2011
2012         trace_ext4_writepage(page);
2013         size = i_size_read(inode);
2014         if (page->index == size >> PAGE_SHIFT)
2015                 len = size & ~PAGE_MASK;
2016         else
2017                 len = PAGE_SIZE;
2018
2019         page_bufs = page_buffers(page);
2020         /*
2021          * We cannot do block allocation or other extent handling in this
2022          * function. If there are buffers needing that, we have to redirty
2023          * the page. But we may reach here when we do a journal commit via
2024          * journal_submit_inode_data_buffers() and in that case we must write
2025          * allocated buffers to achieve data=ordered mode guarantees.
2026          *
2027          * Also, if there is only one buffer per page (the fs block
2028          * size == the page size), if one buffer needs block
2029          * allocation or needs to modify the extent tree to clear the
2030          * unwritten flag, we know that the page can't be written at
2031          * all, so we might as well refuse the write immediately.
2032          * Unfortunately if the block size != page size, we can't as
2033          * easily detect this case using ext4_walk_page_buffers(), but
2034          * for the extremely common case, this is an optimization that
2035          * skips a useless round trip through ext4_bio_write_page().
2036          */
2037         if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2038                                    ext4_bh_delay_or_unwritten)) {
2039                 redirty_page_for_writepage(wbc, page);
2040                 if ((current->flags & PF_MEMALLOC) ||
2041                     (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2042                         /*
2043                          * For memory cleaning there's no point in writing only
2044                          * some buffers. So just bail out. Warn if we came here
2045                          * from direct reclaim.
2046                          */
2047                         WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2048                                                         == PF_MEMALLOC);
2049                         unlock_page(page);
2050                         return 0;
2051                 }
2052                 keep_towrite = true;
2053         }
2054
2055         if (PageChecked(page) && ext4_should_journal_data(inode))
2056                 /*
2057                  * It's mmapped pagecache.  Add buffers and journal it.  There
2058                  * doesn't seem much point in redirtying the page here.
2059                  */
2060                 return __ext4_journalled_writepage(page, len);
2061
2062         ext4_io_submit_init(&io_submit, wbc);
2063         io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2064         if (!io_submit.io_end) {
2065                 redirty_page_for_writepage(wbc, page);
2066                 unlock_page(page);
2067                 return -ENOMEM;
2068         }
2069         ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2070         ext4_io_submit(&io_submit);
2071         /* Drop io_end reference we got from init */
2072         ext4_put_io_end_defer(io_submit.io_end);
2073         return ret;
2074 }
2075
2076 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2077 {
2078         int len;
2079         loff_t size = i_size_read(mpd->inode);
2080         int err;
2081
2082         BUG_ON(page->index != mpd->first_page);
2083         if (page->index == size >> PAGE_SHIFT)
2084                 len = size & ~PAGE_MASK;
2085         else
2086                 len = PAGE_SIZE;
2087         clear_page_dirty_for_io(page);
2088         err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2089         if (!err)
2090                 mpd->wbc->nr_to_write--;
2091         mpd->first_page++;
2092
2093         return err;
2094 }
2095
2096 #define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2097
2098 /*
2099  * mballoc gives us at most this number of blocks...
2100  * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2101  * The rest of mballoc seems to handle chunks up to full group size.
2102  */
2103 #define MAX_WRITEPAGES_EXTENT_LEN 2048
2104
2105 /*
2106  * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2107  *
2108  * @mpd - extent of blocks
2109  * @lblk - logical number of the block in the file
2110  * @bh - buffer head we want to add to the extent
2111  *
2112  * The function is used to collect contig. blocks in the same state. If the
2113  * buffer doesn't require mapping for writeback and we haven't started the
2114  * extent of buffers to map yet, the function returns 'true' immediately - the
2115  * caller can write the buffer right away. Otherwise the function returns true
2116  * if the block has been added to the extent, false if the block couldn't be
2117  * added.
2118  */
2119 static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2120                                    struct buffer_head *bh)
2121 {
2122         struct ext4_map_blocks *map = &mpd->map;
2123
2124         /* Buffer that doesn't need mapping for writeback? */
2125         if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2126             (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2127                 /* So far no extent to map => we write the buffer right away */
2128                 if (map->m_len == 0)
2129                         return true;
2130                 return false;
2131         }
2132
2133         /* First block in the extent? */
2134         if (map->m_len == 0) {
2135                 map->m_lblk = lblk;
2136                 map->m_len = 1;
2137                 map->m_flags = bh->b_state & BH_FLAGS;
2138                 return true;
2139         }
2140
2141         /* Don't go larger than mballoc is willing to allocate */
2142         if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2143                 return false;
2144
2145         /* Can we merge the block to our big extent? */
2146         if (lblk == map->m_lblk + map->m_len &&
2147             (bh->b_state & BH_FLAGS) == map->m_flags) {
2148                 map->m_len++;
2149                 return true;
2150         }
2151         return false;
2152 }
2153
2154 /*
2155  * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2156  *
2157  * @mpd - extent of blocks for mapping
2158  * @head - the first buffer in the page
2159  * @bh - buffer we should start processing from
2160  * @lblk - logical number of the block in the file corresponding to @bh
2161  *
2162  * Walk through page buffers from @bh upto @head (exclusive) and either submit
2163  * the page for IO if all buffers in this page were mapped and there's no
2164  * accumulated extent of buffers to map or add buffers in the page to the
2165  * extent of buffers to map. The function returns 1 if the caller can continue
2166  * by processing the next page, 0 if it should stop adding buffers to the
2167  * extent to map because we cannot extend it anymore. It can also return value
2168  * < 0 in case of error during IO submission.
2169  */
2170 static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2171                                    struct buffer_head *head,
2172                                    struct buffer_head *bh,
2173                                    ext4_lblk_t lblk)
2174 {
2175         struct inode *inode = mpd->inode;
2176         int err;
2177         ext4_lblk_t blocks = (i_size_read(inode) + (1 << inode->i_blkbits) - 1)
2178                                                         >> inode->i_blkbits;
2179
2180         do {
2181                 BUG_ON(buffer_locked(bh));
2182
2183                 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2184                         /* Found extent to map? */
2185                         if (mpd->map.m_len)
2186                                 return 0;
2187                         /* Everything mapped so far and we hit EOF */
2188                         break;
2189                 }
2190         } while (lblk++, (bh = bh->b_this_page) != head);
2191         /* So far everything mapped? Submit the page for IO. */
2192         if (mpd->map.m_len == 0) {
2193                 err = mpage_submit_page(mpd, head->b_page);
2194                 if (err < 0)
2195                         return err;
2196         }
2197         return lblk < blocks;
2198 }
2199
2200 /*
2201  * mpage_map_buffers - update buffers corresponding to changed extent and
2202  *                     submit fully mapped pages for IO
2203  *
2204  * @mpd - description of extent to map, on return next extent to map
2205  *
2206  * Scan buffers corresponding to changed extent (we expect corresponding pages
2207  * to be already locked) and update buffer state according to new extent state.
2208  * We map delalloc buffers to their physical location, clear unwritten bits,
2209  * and mark buffers as uninit when we perform writes to unwritten extents
2210  * and do extent conversion after IO is finished. If the last page is not fully
2211  * mapped, we update @map to the next extent in the last page that needs
2212  * mapping. Otherwise we submit the page for IO.
2213  */
2214 static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2215 {
2216         struct pagevec pvec;
2217         int nr_pages, i;
2218         struct inode *inode = mpd->inode;
2219         struct buffer_head *head, *bh;
2220         int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2221         pgoff_t start, end;
2222         ext4_lblk_t lblk;
2223         sector_t pblock;
2224         int err;
2225
2226         start = mpd->map.m_lblk >> bpp_bits;
2227         end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2228         lblk = start << bpp_bits;
2229         pblock = mpd->map.m_pblk;
2230
2231         pagevec_init(&pvec, 0);
2232         while (start <= end) {
2233                 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, start,
2234                                           PAGEVEC_SIZE);
2235                 if (nr_pages == 0)
2236                         break;
2237                 for (i = 0; i < nr_pages; i++) {
2238                         struct page *page = pvec.pages[i];
2239
2240                         if (page->index > end)
2241                                 break;
2242                         /* Up to 'end' pages must be contiguous */
2243                         BUG_ON(page->index != start);
2244                         bh = head = page_buffers(page);
2245                         do {
2246                                 if (lblk < mpd->map.m_lblk)
2247                                         continue;
2248                                 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2249                                         /*
2250                                          * Buffer after end of mapped extent.
2251                                          * Find next buffer in the page to map.
2252                                          */
2253                                         mpd->map.m_len = 0;
2254                                         mpd->map.m_flags = 0;
2255                                         /*
2256                                          * FIXME: If dioread_nolock supports
2257                                          * blocksize < pagesize, we need to make
2258                                          * sure we add size mapped so far to
2259                                          * io_end->size as the following call
2260                                          * can submit the page for IO.
2261                                          */
2262                                         err = mpage_process_page_bufs(mpd, head,
2263                                                                       bh, lblk);
2264                                         pagevec_release(&pvec);
2265                                         if (err > 0)
2266                                                 err = 0;
2267                                         return err;
2268                                 }
2269                                 if (buffer_delay(bh)) {
2270                                         clear_buffer_delay(bh);
2271                                         bh->b_blocknr = pblock++;
2272                                 }
2273                                 clear_buffer_unwritten(bh);
2274                         } while (lblk++, (bh = bh->b_this_page) != head);
2275
2276                         /*
2277                          * FIXME: This is going to break if dioread_nolock
2278                          * supports blocksize < pagesize as we will try to
2279                          * convert potentially unmapped parts of inode.
2280                          */
2281                         mpd->io_submit.io_end->size += PAGE_SIZE;
2282                         /* Page fully mapped - let IO run! */
2283                         err = mpage_submit_page(mpd, page);
2284                         if (err < 0) {
2285                                 pagevec_release(&pvec);
2286                                 return err;
2287                         }
2288                         start++;
2289                 }
2290                 pagevec_release(&pvec);
2291         }
2292         /* Extent fully mapped and matches with page boundary. We are done. */
2293         mpd->map.m_len = 0;
2294         mpd->map.m_flags = 0;
2295         return 0;
2296 }
2297
2298 static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2299 {
2300         struct inode *inode = mpd->inode;
2301         struct ext4_map_blocks *map = &mpd->map;
2302         int get_blocks_flags;
2303         int err, dioread_nolock;
2304
2305         trace_ext4_da_write_pages_extent(inode, map);
2306         /*
2307          * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2308          * to convert an unwritten extent to be initialized (in the case
2309          * where we have written into one or more preallocated blocks).  It is
2310          * possible that we're going to need more metadata blocks than
2311          * previously reserved. However we must not fail because we're in
2312          * writeback and there is nothing we can do about it so it might result
2313          * in data loss.  So use reserved blocks to allocate metadata if
2314          * possible.
2315          *
2316          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2317          * the blocks in question are delalloc blocks.  This indicates
2318          * that the blocks and quotas has already been checked when
2319          * the data was copied into the page cache.
2320          */
2321         get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2322                            EXT4_GET_BLOCKS_METADATA_NOFAIL;
2323         dioread_nolock = ext4_should_dioread_nolock(inode);
2324         if (dioread_nolock)
2325                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2326         if (map->m_flags & (1 << BH_Delay))
2327                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2328
2329         err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2330         if (err < 0)
2331                 return err;
2332         if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2333                 if (!mpd->io_submit.io_end->handle &&
2334                     ext4_handle_valid(handle)) {
2335                         mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2336                         handle->h_rsv_handle = NULL;
2337                 }
2338                 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2339         }
2340
2341         BUG_ON(map->m_len == 0);
2342         if (map->m_flags & EXT4_MAP_NEW) {
2343                 struct block_device *bdev = inode->i_sb->s_bdev;
2344                 int i;
2345
2346                 for (i = 0; i < map->m_len; i++)
2347                         unmap_underlying_metadata(bdev, map->m_pblk + i);
2348         }
2349         return 0;
2350 }
2351
2352 /*
2353  * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2354  *                               mpd->len and submit pages underlying it for IO
2355  *
2356  * @handle - handle for journal operations
2357  * @mpd - extent to map
2358  * @give_up_on_write - we set this to true iff there is a fatal error and there
2359  *                     is no hope of writing the data. The caller should discard
2360  *                     dirty pages to avoid infinite loops.
2361  *
2362  * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2363  * delayed, blocks are allocated, if it is unwritten, we may need to convert
2364  * them to initialized or split the described range from larger unwritten
2365  * extent. Note that we need not map all the described range since allocation
2366  * can return less blocks or the range is covered by more unwritten extents. We
2367  * cannot map more because we are limited by reserved transaction credits. On
2368  * the other hand we always make sure that the last touched page is fully
2369  * mapped so that it can be written out (and thus forward progress is
2370  * guaranteed). After mapping we submit all mapped pages for IO.
2371  */
2372 static int mpage_map_and_submit_extent(handle_t *handle,
2373                                        struct mpage_da_data *mpd,
2374                                        bool *give_up_on_write)
2375 {
2376         struct inode *inode = mpd->inode;
2377         struct ext4_map_blocks *map = &mpd->map;
2378         int err;
2379         loff_t disksize;
2380         int progress = 0;
2381
2382         mpd->io_submit.io_end->offset =
2383                                 ((loff_t)map->m_lblk) << inode->i_blkbits;
2384         do {
2385                 err = mpage_map_one_extent(handle, mpd);
2386                 if (err < 0) {
2387                         struct super_block *sb = inode->i_sb;
2388
2389                         if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2390                                 goto invalidate_dirty_pages;
2391                         /*
2392                          * Let the uper layers retry transient errors.
2393                          * In the case of ENOSPC, if ext4_count_free_blocks()
2394                          * is non-zero, a commit should free up blocks.
2395                          */
2396                         if ((err == -ENOMEM) ||
2397                             (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2398                                 if (progress)
2399                                         goto update_disksize;
2400                                 return err;
2401                         }
2402                         ext4_msg(sb, KERN_CRIT,
2403                                  "Delayed block allocation failed for "
2404                                  "inode %lu at logical offset %llu with"
2405                                  " max blocks %u with error %d",
2406                                  inode->i_ino,
2407                                  (unsigned long long)map->m_lblk,
2408                                  (unsigned)map->m_len, -err);
2409                         ext4_msg(sb, KERN_CRIT,
2410                                  "This should not happen!! Data will "
2411                                  "be lost\n");
2412                         if (err == -ENOSPC)
2413                                 ext4_print_free_blocks(inode);
2414                 invalidate_dirty_pages:
2415                         *give_up_on_write = true;
2416                         return err;
2417                 }
2418                 progress = 1;
2419                 /*
2420                  * Update buffer state, submit mapped pages, and get us new
2421                  * extent to map
2422                  */
2423                 err = mpage_map_and_submit_buffers(mpd);
2424                 if (err < 0)
2425                         goto update_disksize;
2426         } while (map->m_len);
2427
2428 update_disksize:
2429         /*
2430          * Update on-disk size after IO is submitted.  Races with
2431          * truncate are avoided by checking i_size under i_data_sem.
2432          */
2433         disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2434         if (disksize > EXT4_I(inode)->i_disksize) {
2435                 int err2;
2436                 loff_t i_size;
2437
2438                 down_write(&EXT4_I(inode)->i_data_sem);
2439                 i_size = i_size_read(inode);
2440                 if (disksize > i_size)
2441                         disksize = i_size;
2442                 if (disksize > EXT4_I(inode)->i_disksize)
2443                         EXT4_I(inode)->i_disksize = disksize;
2444                 err2 = ext4_mark_inode_dirty(handle, inode);
2445                 up_write(&EXT4_I(inode)->i_data_sem);
2446                 if (err2)
2447                         ext4_error(inode->i_sb,
2448                                    "Failed to mark inode %lu dirty",
2449                                    inode->i_ino);
2450                 if (!err)
2451                         err = err2;
2452         }
2453         return err;
2454 }
2455
2456 /*
2457  * Calculate the total number of credits to reserve for one writepages
2458  * iteration. This is called from ext4_writepages(). We map an extent of
2459  * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2460  * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2461  * bpp - 1 blocks in bpp different extents.
2462  */
2463 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2464 {
2465         int bpp = ext4_journal_blocks_per_page(inode);
2466
2467         return ext4_meta_trans_blocks(inode,
2468                                 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2469 }
2470
2471 /*
2472  * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2473  *                               and underlying extent to map
2474  *
2475  * @mpd - where to look for pages
2476  *
2477  * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2478  * IO immediately. When we find a page which isn't mapped we start accumulating
2479  * extent of buffers underlying these pages that needs mapping (formed by
2480  * either delayed or unwritten buffers). We also lock the pages containing
2481  * these buffers. The extent found is returned in @mpd structure (starting at
2482  * mpd->lblk with length mpd->len blocks).
2483  *
2484  * Note that this function can attach bios to one io_end structure which are
2485  * neither logically nor physically contiguous. Although it may seem as an
2486  * unnecessary complication, it is actually inevitable in blocksize < pagesize
2487  * case as we need to track IO to all buffers underlying a page in one io_end.
2488  */
2489 static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2490 {
2491         struct address_space *mapping = mpd->inode->i_mapping;
2492         struct pagevec pvec;
2493         unsigned int nr_pages;
2494         long left = mpd->wbc->nr_to_write;
2495         pgoff_t index = mpd->first_page;
2496         pgoff_t end = mpd->last_page;
2497         int tag;
2498         int i, err = 0;
2499         int blkbits = mpd->inode->i_blkbits;
2500         ext4_lblk_t lblk;
2501         struct buffer_head *head;
2502
2503         if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2504                 tag = PAGECACHE_TAG_TOWRITE;
2505         else
2506                 tag = PAGECACHE_TAG_DIRTY;
2507
2508         pagevec_init(&pvec, 0);
2509         mpd->map.m_len = 0;
2510         mpd->next_page = index;
2511         while (index <= end) {
2512                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2513                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2514                 if (nr_pages == 0)
2515                         goto out;
2516
2517                 for (i = 0; i < nr_pages; i++) {
2518                         struct page *page = pvec.pages[i];
2519
2520                         /*
2521                          * At this point, the page may be truncated or
2522                          * invalidated (changing page->mapping to NULL), or
2523                          * even swizzled back from swapper_space to tmpfs file
2524                          * mapping. However, page->index will not change
2525                          * because we have a reference on the page.
2526                          */
2527                         if (page->index > end)
2528                                 goto out;
2529
2530                         /*
2531                          * Accumulated enough dirty pages? This doesn't apply
2532                          * to WB_SYNC_ALL mode. For integrity sync we have to
2533                          * keep going because someone may be concurrently
2534                          * dirtying pages, and we might have synced a lot of
2535                          * newly appeared dirty pages, but have not synced all
2536                          * of the old dirty pages.
2537                          */
2538                         if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2539                                 goto out;
2540
2541                         /* If we can't merge this page, we are done. */
2542                         if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2543                                 goto out;
2544
2545                         lock_page(page);
2546                         /*
2547                          * If the page is no longer dirty, or its mapping no
2548                          * longer corresponds to inode we are writing (which
2549                          * means it has been truncated or invalidated), or the
2550                          * page is already under writeback and we are not doing
2551                          * a data integrity writeback, skip the page
2552                          */
2553                         if (!PageDirty(page) ||
2554                             (PageWriteback(page) &&
2555                              (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2556                             unlikely(page->mapping != mapping)) {
2557                                 unlock_page(page);
2558                                 continue;
2559                         }
2560
2561                         wait_on_page_writeback(page);
2562                         BUG_ON(PageWriteback(page));
2563
2564                         if (mpd->map.m_len == 0)
2565                                 mpd->first_page = page->index;
2566                         mpd->next_page = page->index + 1;
2567                         /* Add all dirty buffers to mpd */
2568                         lblk = ((ext4_lblk_t)page->index) <<
2569                                 (PAGE_SHIFT - blkbits);
2570                         head = page_buffers(page);
2571                         err = mpage_process_page_bufs(mpd, head, head, lblk);
2572                         if (err <= 0)
2573                                 goto out;
2574                         err = 0;
2575                         left--;
2576                 }
2577                 pagevec_release(&pvec);
2578                 cond_resched();
2579         }
2580         return 0;
2581 out:
2582         pagevec_release(&pvec);
2583         return err;
2584 }
2585
2586 static int __writepage(struct page *page, struct writeback_control *wbc,
2587                        void *data)
2588 {
2589         struct address_space *mapping = data;
2590         int ret = ext4_writepage(page, wbc);
2591         mapping_set_error(mapping, ret);
2592         return ret;
2593 }
2594
2595 static int ext4_writepages(struct address_space *mapping,
2596                            struct writeback_control *wbc)
2597 {
2598         pgoff_t writeback_index = 0;
2599         long nr_to_write = wbc->nr_to_write;
2600         int range_whole = 0;
2601         int cycled = 1;
2602         handle_t *handle = NULL;
2603         struct mpage_da_data mpd;
2604         struct inode *inode = mapping->host;
2605         int needed_blocks, rsv_blocks = 0, ret = 0;
2606         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2607         bool done;
2608         struct blk_plug plug;
2609         bool give_up_on_write = false;
2610
2611         trace_ext4_writepages(inode, wbc);
2612
2613         if (dax_mapping(mapping))
2614                 return dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev,
2615                                                    wbc);
2616
2617         /*
2618          * No pages to write? This is mainly a kludge to avoid starting
2619          * a transaction for special inodes like journal inode on last iput()
2620          * because that could violate lock ordering on umount
2621          */
2622         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2623                 goto out_writepages;
2624
2625         if (ext4_should_journal_data(inode)) {
2626                 struct blk_plug plug;
2627
2628                 blk_start_plug(&plug);
2629                 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2630                 blk_finish_plug(&plug);
2631                 goto out_writepages;
2632         }
2633
2634         /*
2635          * If the filesystem has aborted, it is read-only, so return
2636          * right away instead of dumping stack traces later on that
2637          * will obscure the real source of the problem.  We test
2638          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2639          * the latter could be true if the filesystem is mounted
2640          * read-only, and in that case, ext4_writepages should
2641          * *never* be called, so if that ever happens, we would want
2642          * the stack trace.
2643          */
2644         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2645                 ret = -EROFS;
2646                 goto out_writepages;
2647         }
2648
2649         if (ext4_should_dioread_nolock(inode)) {
2650                 /*
2651                  * We may need to convert up to one extent per block in
2652                  * the page and we may dirty the inode.
2653                  */
2654                 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2655         }
2656
2657         /*
2658          * If we have inline data and arrive here, it means that
2659          * we will soon create the block for the 1st page, so
2660          * we'd better clear the inline data here.
2661          */
2662         if (ext4_has_inline_data(inode)) {
2663                 /* Just inode will be modified... */
2664                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2665                 if (IS_ERR(handle)) {
2666                         ret = PTR_ERR(handle);
2667                         goto out_writepages;
2668                 }
2669                 BUG_ON(ext4_test_inode_state(inode,
2670                                 EXT4_STATE_MAY_INLINE_DATA));
2671                 ext4_destroy_inline_data(handle, inode);
2672                 ext4_journal_stop(handle);
2673         }
2674
2675         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2676                 range_whole = 1;
2677
2678         if (wbc->range_cyclic) {
2679                 writeback_index = mapping->writeback_index;
2680                 if (writeback_index)
2681                         cycled = 0;
2682                 mpd.first_page = writeback_index;
2683                 mpd.last_page = -1;
2684         } else {
2685                 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2686                 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2687         }
2688
2689         mpd.inode = inode;
2690         mpd.wbc = wbc;
2691         ext4_io_submit_init(&mpd.io_submit, wbc);
2692 retry:
2693         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2694                 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2695         done = false;
2696         blk_start_plug(&plug);
2697         while (!done && mpd.first_page <= mpd.last_page) {
2698                 /* For each extent of pages we use new io_end */
2699                 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2700                 if (!mpd.io_submit.io_end) {
2701                         ret = -ENOMEM;
2702                         break;
2703                 }
2704
2705                 /*
2706                  * We have two constraints: We find one extent to map and we
2707                  * must always write out whole page (makes a difference when
2708                  * blocksize < pagesize) so that we don't block on IO when we
2709                  * try to write out the rest of the page. Journalled mode is
2710                  * not supported by delalloc.
2711                  */
2712                 BUG_ON(ext4_should_journal_data(inode));
2713                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2714
2715                 /* start a new transaction */
2716                 handle = ext4_journal_start_with_reserve(inode,
2717                                 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2718                 if (IS_ERR(handle)) {
2719                         ret = PTR_ERR(handle);
2720                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2721                                "%ld pages, ino %lu; err %d", __func__,
2722                                 wbc->nr_to_write, inode->i_ino, ret);
2723                         /* Release allocated io_end */
2724                         ext4_put_io_end(mpd.io_submit.io_end);
2725                         break;
2726                 }
2727
2728                 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2729                 ret = mpage_prepare_extent_to_map(&mpd);
2730                 if (!ret) {
2731                         if (mpd.map.m_len)
2732                                 ret = mpage_map_and_submit_extent(handle, &mpd,
2733                                         &give_up_on_write);
2734                         else {
2735                                 /*
2736                                  * We scanned the whole range (or exhausted
2737                                  * nr_to_write), submitted what was mapped and
2738                                  * didn't find anything needing mapping. We are
2739                                  * done.
2740                                  */
2741                                 done = true;
2742                         }
2743                 }
2744                 ext4_journal_stop(handle);
2745                 /* Submit prepared bio */
2746                 ext4_io_submit(&mpd.io_submit);
2747                 /* Unlock pages we didn't use */
2748                 mpage_release_unused_pages(&mpd, give_up_on_write);
2749                 /* Drop our io_end reference we got from init */
2750                 ext4_put_io_end(mpd.io_submit.io_end);
2751
2752                 if (ret == -ENOSPC && sbi->s_journal) {
2753                         /*
2754                          * Commit the transaction which would
2755                          * free blocks released in the transaction
2756                          * and try again
2757                          */
2758                         jbd2_journal_force_commit_nested(sbi->s_journal);
2759                         ret = 0;
2760                         continue;
2761                 }
2762                 /* Fatal error - ENOMEM, EIO... */
2763                 if (ret)
2764                         break;
2765         }
2766         blk_finish_plug(&plug);
2767         if (!ret && !cycled && wbc->nr_to_write > 0) {
2768                 cycled = 1;
2769                 mpd.last_page = writeback_index - 1;
2770                 mpd.first_page = 0;
2771                 goto retry;
2772         }
2773
2774         /* Update index */
2775         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2776                 /*
2777                  * Set the writeback_index so that range_cyclic
2778                  * mode will write it back later
2779                  */
2780                 mapping->writeback_index = mpd.first_page;
2781
2782 out_writepages:
2783         trace_ext4_writepages_result(inode, wbc, ret,
2784                                      nr_to_write - wbc->nr_to_write);
2785         return ret;
2786 }
2787
2788 static int ext4_nonda_switch(struct super_block *sb)
2789 {
2790         s64 free_clusters, dirty_clusters;
2791         struct ext4_sb_info *sbi = EXT4_SB(sb);
2792
2793         /*
2794          * switch to non delalloc mode if we are running low
2795          * on free block. The free block accounting via percpu
2796          * counters can get slightly wrong with percpu_counter_batch getting
2797          * accumulated on each CPU without updating global counters
2798          * Delalloc need an accurate free block accounting. So switch
2799          * to non delalloc when we are near to error range.
2800          */
2801         free_clusters =
2802                 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2803         dirty_clusters =
2804                 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2805         /*
2806          * Start pushing delalloc when 1/2 of free blocks are dirty.
2807          */
2808         if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2809                 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2810
2811         if (2 * free_clusters < 3 * dirty_clusters ||
2812             free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2813                 /*
2814                  * free block count is less than 150% of dirty blocks
2815                  * or free blocks is less than watermark
2816                  */
2817                 return 1;
2818         }
2819         return 0;
2820 }
2821
2822 /* We always reserve for an inode update; the superblock could be there too */
2823 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2824 {
2825         if (likely(ext4_has_feature_large_file(inode->i_sb)))
2826                 return 1;
2827
2828         if (pos + len <= 0x7fffffffULL)
2829                 return 1;
2830
2831         /* We might need to update the superblock to set LARGE_FILE */
2832         return 2;
2833 }
2834
2835 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2836                                loff_t pos, unsigned len, unsigned flags,
2837                                struct page **pagep, void **fsdata)
2838 {
2839         int ret, retries = 0;
2840         struct page *page;
2841         pgoff_t index;
2842         struct inode *inode = mapping->host;
2843         handle_t *handle;
2844
2845         index = pos >> PAGE_SHIFT;
2846
2847         if (ext4_nonda_switch(inode->i_sb)) {
2848                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2849                 return ext4_write_begin(file, mapping, pos,
2850                                         len, flags, pagep, fsdata);
2851         }
2852         *fsdata = (void *)0;
2853         trace_ext4_da_write_begin(inode, pos, len, flags);
2854
2855         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2856                 ret = ext4_da_write_inline_data_begin(mapping, inode,
2857                                                       pos, len, flags,
2858                                                       pagep, fsdata);
2859                 if (ret < 0)
2860                         return ret;
2861                 if (ret == 1)
2862                         return 0;
2863         }
2864
2865         /*
2866          * grab_cache_page_write_begin() can take a long time if the
2867          * system is thrashing due to memory pressure, or if the page
2868          * is being written back.  So grab it first before we start
2869          * the transaction handle.  This also allows us to allocate
2870          * the page (if needed) without using GFP_NOFS.
2871          */
2872 retry_grab:
2873         page = grab_cache_page_write_begin(mapping, index, flags);
2874         if (!page)
2875                 return -ENOMEM;
2876         unlock_page(page);
2877
2878         /*
2879          * With delayed allocation, we don't log the i_disksize update
2880          * if there is delayed block allocation. But we still need
2881          * to journalling the i_disksize update if writes to the end
2882          * of file which has an already mapped buffer.
2883          */
2884 retry_journal:
2885         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2886                                 ext4_da_write_credits(inode, pos, len));
2887         if (IS_ERR(handle)) {
2888                 put_page(page);
2889                 return PTR_ERR(handle);
2890         }
2891
2892         lock_page(page);
2893         if (page->mapping != mapping) {
2894                 /* The page got truncated from under us */
2895                 unlock_page(page);
2896                 put_page(page);
2897                 ext4_journal_stop(handle);
2898                 goto retry_grab;
2899         }
2900         /* In case writeback began while the page was unlocked */
2901         wait_for_stable_page(page);
2902
2903 #ifdef CONFIG_EXT4_FS_ENCRYPTION
2904         ret = ext4_block_write_begin(page, pos, len,
2905                                      ext4_da_get_block_prep);
2906 #else
2907         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2908 #endif
2909         if (ret < 0) {
2910                 unlock_page(page);
2911                 ext4_journal_stop(handle);
2912                 /*
2913                  * block_write_begin may have instantiated a few blocks
2914                  * outside i_size.  Trim these off again. Don't need
2915                  * i_size_read because we hold i_mutex.
2916                  */
2917                 if (pos + len > inode->i_size)
2918                         ext4_truncate_failed_write(inode);
2919
2920                 if (ret == -ENOSPC &&
2921                     ext4_should_retry_alloc(inode->i_sb, &retries))
2922                         goto retry_journal;
2923
2924                 put_page(page);
2925                 return ret;
2926         }
2927
2928         *pagep = page;
2929         return ret;
2930 }
2931
2932 /*
2933  * Check if we should update i_disksize
2934  * when write to the end of file but not require block allocation
2935  */
2936 static int ext4_da_should_update_i_disksize(struct page *page,
2937                                             unsigned long offset)
2938 {
2939         struct buffer_head *bh;
2940         struct inode *inode = page->mapping->host;
2941         unsigned int idx;
2942         int i;
2943
2944         bh = page_buffers(page);
2945         idx = offset >> inode->i_blkbits;
2946
2947         for (i = 0; i < idx; i++)
2948                 bh = bh->b_this_page;
2949
2950         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2951                 return 0;
2952         return 1;
2953 }
2954
2955 static int ext4_da_write_end(struct file *file,
2956                              struct address_space *mapping,
2957                              loff_t pos, unsigned len, unsigned copied,
2958                              struct page *page, void *fsdata)
2959 {
2960         struct inode *inode = mapping->host;
2961         int ret = 0, ret2;
2962         handle_t *handle = ext4_journal_current_handle();
2963         loff_t new_i_size;
2964         unsigned long start, end;
2965         int write_mode = (int)(unsigned long)fsdata;
2966
2967         if (write_mode == FALL_BACK_TO_NONDELALLOC)
2968                 return ext4_write_end(file, mapping, pos,
2969                                       len, copied, page, fsdata);
2970
2971         trace_ext4_da_write_end(inode, pos, len, copied);
2972         start = pos & (PAGE_SIZE - 1);
2973         end = start + copied - 1;
2974
2975         /*
2976          * generic_write_end() will run mark_inode_dirty() if i_size
2977          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2978          * into that.
2979          */
2980         new_i_size = pos + copied;
2981         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
2982                 if (ext4_has_inline_data(inode) ||
2983                     ext4_da_should_update_i_disksize(page, end)) {
2984                         ext4_update_i_disksize(inode, new_i_size);
2985                         /* We need to mark inode dirty even if
2986                          * new_i_size is less that inode->i_size
2987                          * bu greater than i_disksize.(hint delalloc)
2988                          */
2989                         ext4_mark_inode_dirty(handle, inode);
2990                 }
2991         }
2992
2993         if (write_mode != CONVERT_INLINE_DATA &&
2994             ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
2995             ext4_has_inline_data(inode))
2996                 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
2997                                                      page);
2998         else
2999                 ret2 = generic_write_end(file, mapping, pos, len, copied,
3000                                                         page, fsdata);
3001
3002         copied = ret2;
3003         if (ret2 < 0)
3004                 ret = ret2;
3005         ret2 = ext4_journal_stop(handle);
3006         if (!ret)
3007                 ret = ret2;
3008
3009         return ret ? ret : copied;
3010 }
3011
3012 static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3013                                    unsigned int length)
3014 {
3015         /*
3016          * Drop reserved blocks
3017          */
3018         BUG_ON(!PageLocked(page));
3019         if (!page_has_buffers(page))
3020                 goto out;
3021
3022         ext4_da_page_release_reservation(page, offset, length);
3023
3024 out:
3025         ext4_invalidatepage(page, offset, length);
3026
3027         return;
3028 }
3029
3030 /*
3031  * Force all delayed allocation blocks to be allocated for a given inode.
3032  */
3033 int ext4_alloc_da_blocks(struct inode *inode)
3034 {
3035         trace_ext4_alloc_da_blocks(inode);
3036
3037         if (!EXT4_I(inode)->i_reserved_data_blocks)
3038                 return 0;
3039
3040         /*
3041          * We do something simple for now.  The filemap_flush() will
3042          * also start triggering a write of the data blocks, which is
3043          * not strictly speaking necessary (and for users of
3044          * laptop_mode, not even desirable).  However, to do otherwise
3045          * would require replicating code paths in:
3046          *
3047          * ext4_writepages() ->
3048          *    write_cache_pages() ---> (via passed in callback function)
3049          *        __mpage_da_writepage() -->
3050          *           mpage_add_bh_to_extent()
3051          *           mpage_da_map_blocks()
3052          *
3053          * The problem is that write_cache_pages(), located in
3054          * mm/page-writeback.c, marks pages clean in preparation for
3055          * doing I/O, which is not desirable if we're not planning on
3056          * doing I/O at all.
3057          *
3058          * We could call write_cache_pages(), and then redirty all of
3059          * the pages by calling redirty_page_for_writepage() but that
3060          * would be ugly in the extreme.  So instead we would need to
3061          * replicate parts of the code in the above functions,
3062          * simplifying them because we wouldn't actually intend to
3063          * write out the pages, but rather only collect contiguous
3064          * logical block extents, call the multi-block allocator, and
3065          * then update the buffer heads with the block allocations.
3066          *
3067          * For now, though, we'll cheat by calling filemap_flush(),
3068          * which will map the blocks, and start the I/O, but not
3069          * actually wait for the I/O to complete.
3070          */
3071         return filemap_flush(inode->i_mapping);
3072 }
3073
3074 /*
3075  * bmap() is special.  It gets used by applications such as lilo and by
3076  * the swapper to find the on-disk block of a specific piece of data.
3077  *
3078  * Naturally, this is dangerous if the block concerned is still in the
3079  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3080  * filesystem and enables swap, then they may get a nasty shock when the
3081  * data getting swapped to that swapfile suddenly gets overwritten by
3082  * the original zero's written out previously to the journal and
3083  * awaiting writeback in the kernel's buffer cache.
3084  *
3085  * So, if we see any bmap calls here on a modified, data-journaled file,
3086  * take extra steps to flush any blocks which might be in the cache.
3087  */
3088 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3089 {
3090         struct inode *inode = mapping->host;
3091         journal_t *journal;
3092         int err;
3093
3094         /*
3095          * We can get here for an inline file via the FIBMAP ioctl
3096          */
3097         if (ext4_has_inline_data(inode))
3098                 return 0;
3099
3100         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3101                         test_opt(inode->i_sb, DELALLOC)) {
3102                 /*
3103                  * With delalloc we want to sync the file
3104                  * so that we can make sure we allocate
3105                  * blocks for file
3106                  */
3107                 filemap_write_and_wait(mapping);
3108         }
3109
3110         if (EXT4_JOURNAL(inode) &&
3111             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3112                 /*
3113                  * This is a REALLY heavyweight approach, but the use of
3114                  * bmap on dirty files is expected to be extremely rare:
3115                  * only if we run lilo or swapon on a freshly made file
3116                  * do we expect this to happen.
3117                  *
3118                  * (bmap requires CAP_SYS_RAWIO so this does not
3119                  * represent an unprivileged user DOS attack --- we'd be
3120                  * in trouble if mortal users could trigger this path at
3121                  * will.)
3122                  *
3123                  * NB. EXT4_STATE_JDATA is not set on files other than
3124                  * regular files.  If somebody wants to bmap a directory
3125                  * or symlink and gets confused because the buffer
3126                  * hasn't yet been flushed to disk, they deserve
3127                  * everything they get.
3128                  */
3129
3130                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3131                 journal = EXT4_JOURNAL(inode);
3132                 jbd2_journal_lock_updates(journal);
3133                 err = jbd2_journal_flush(journal);
3134                 jbd2_journal_unlock_updates(journal);
3135
3136                 if (err)
3137                         return 0;
3138         }
3139
3140         return generic_block_bmap(mapping, block, ext4_get_block);
3141 }
3142
3143 static int ext4_readpage(struct file *file, struct page *page)
3144 {
3145         int ret = -EAGAIN;
3146         struct inode *inode = page->mapping->host;
3147
3148         trace_ext4_readpage(page);
3149
3150         if (ext4_has_inline_data(inode))
3151                 ret = ext4_readpage_inline(inode, page);
3152
3153         if (ret == -EAGAIN)
3154                 return ext4_mpage_readpages(page->mapping, NULL, page, 1);
3155
3156         return ret;
3157 }
3158
3159 static int
3160 ext4_readpages(struct file *file, struct address_space *mapping,
3161                 struct list_head *pages, unsigned nr_pages)
3162 {
3163         struct inode *inode = mapping->host;
3164
3165         /* If the file has inline data, no need to do readpages. */
3166         if (ext4_has_inline_data(inode))
3167                 return 0;
3168
3169         return ext4_mpage_readpages(mapping, pages, NULL, nr_pages);
3170 }
3171
3172 static void ext4_invalidatepage(struct page *page, unsigned int offset,
3173                                 unsigned int length)
3174 {
3175         trace_ext4_invalidatepage(page, offset, length);
3176
3177         /* No journalling happens on data buffers when this function is used */
3178         WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
3179
3180         block_invalidatepage(page, offset, length);
3181 }
3182
3183 static int __ext4_journalled_invalidatepage(struct page *page,
3184                                             unsigned int offset,
3185                                             unsigned int length)
3186 {
3187         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3188
3189         trace_ext4_journalled_invalidatepage(page, offset, length);
3190
3191         /*
3192          * If it's a full truncate we just forget about the pending dirtying
3193          */
3194         if (offset == 0 && length == PAGE_SIZE)
3195                 ClearPageChecked(page);
3196
3197         return jbd2_journal_invalidatepage(journal, page, offset, length);
3198 }
3199
3200 /* Wrapper for aops... */
3201 static void ext4_journalled_invalidatepage(struct page *page,
3202                                            unsigned int offset,
3203                                            unsigned int length)
3204 {
3205         WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
3206 }
3207
3208 static int ext4_releasepage(struct page *page, gfp_t wait)
3209 {
3210         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3211
3212         trace_ext4_releasepage(page);
3213
3214         /* Page has dirty journalled data -> cannot release */
3215         if (PageChecked(page))
3216                 return 0;
3217         if (journal)
3218                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3219         else
3220                 return try_to_free_buffers(page);
3221 }
3222
3223 #ifdef CONFIG_FS_DAX
3224 int ext4_dax_mmap_get_block(struct inode *inode, sector_t iblock,
3225                             struct buffer_head *bh_result, int create)
3226 {
3227         int ret, err;
3228         int credits;
3229         struct ext4_map_blocks map;
3230         handle_t *handle = NULL;
3231         int flags = 0;
3232
3233         ext4_debug("ext4_dax_mmap_get_block: inode %lu, create flag %d\n",
3234                    inode->i_ino, create);
3235         map.m_lblk = iblock;
3236         map.m_len = bh_result->b_size >> inode->i_blkbits;
3237         credits = ext4_chunk_trans_blocks(inode, map.m_len);
3238         if (create) {
3239                 flags |= EXT4_GET_BLOCKS_PRE_IO | EXT4_GET_BLOCKS_CREATE_ZERO;
3240                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, credits);
3241                 if (IS_ERR(handle)) {
3242                         ret = PTR_ERR(handle);
3243                         return ret;
3244                 }
3245         }
3246
3247         ret = ext4_map_blocks(handle, inode, &map, flags);
3248         if (create) {
3249                 err = ext4_journal_stop(handle);
3250                 if (ret >= 0 && err < 0)
3251                         ret = err;
3252         }
3253         if (ret <= 0)
3254                 goto out;
3255         if (map.m_flags & EXT4_MAP_UNWRITTEN) {
3256                 int err2;
3257
3258                 /*
3259                  * We are protected by i_mmap_sem so we know block cannot go
3260                  * away from under us even though we dropped i_data_sem.
3261                  * Convert extent to written and write zeros there.
3262                  *
3263                  * Note: We may get here even when create == 0.
3264                  */
3265                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, credits);
3266                 if (IS_ERR(handle)) {
3267                         ret = PTR_ERR(handle);
3268                         goto out;
3269                 }
3270
3271                 err = ext4_map_blocks(handle, inode, &map,
3272                       EXT4_GET_BLOCKS_CONVERT | EXT4_GET_BLOCKS_CREATE_ZERO);
3273                 if (err < 0)
3274                         ret = err;
3275                 err2 = ext4_journal_stop(handle);
3276                 if (err2 < 0 && ret > 0)
3277                         ret = err2;
3278         }
3279 out:
3280         WARN_ON_ONCE(ret == 0 && create);
3281         if (ret > 0) {
3282                 map_bh(bh_result, inode->i_sb, map.m_pblk);
3283                 /*
3284                  * At least for now we have to clear BH_New so that DAX code
3285                  * doesn't attempt to zero blocks again in a racy way.
3286                  */
3287                 map.m_flags &= ~EXT4_MAP_NEW;
3288                 ext4_update_bh_state(bh_result, map.m_flags);
3289                 bh_result->b_size = map.m_len << inode->i_blkbits;
3290                 ret = 0;
3291         }
3292         return ret;
3293 }
3294 #endif
3295
3296 static int ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3297                             ssize_t size, void *private)
3298 {
3299         ext4_io_end_t *io_end = private;
3300
3301         /* if not async direct IO just return */
3302         if (!io_end)
3303                 return 0;
3304
3305         ext_debug("ext4_end_io_dio(): io_end 0x%p "
3306                   "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
3307                   io_end, io_end->inode->i_ino, iocb, offset, size);
3308
3309         /*
3310          * Error during AIO DIO. We cannot convert unwritten extents as the
3311          * data was not written. Just clear the unwritten flag and drop io_end.
3312          */
3313         if (size <= 0) {
3314                 ext4_clear_io_unwritten_flag(io_end);
3315                 size = 0;
3316         }
3317         io_end->offset = offset;
3318         io_end->size = size;
3319         ext4_put_io_end(io_end);
3320
3321         return 0;
3322 }
3323
3324 /*
3325  * For ext4 extent files, ext4 will do direct-io write to holes,
3326  * preallocated extents, and those write extend the file, no need to
3327  * fall back to buffered IO.
3328  *
3329  * For holes, we fallocate those blocks, mark them as unwritten
3330  * If those blocks were preallocated, we mark sure they are split, but
3331  * still keep the range to write as unwritten.
3332  *
3333  * The unwritten extents will be converted to written when DIO is completed.
3334  * For async direct IO, since the IO may still pending when return, we
3335  * set up an end_io call back function, which will do the conversion
3336  * when async direct IO completed.
3337  *
3338  * If the O_DIRECT write will extend the file then add this inode to the
3339  * orphan list.  So recovery will truncate it back to the original size
3340  * if the machine crashes during the write.
3341  *
3342  */
3343 static ssize_t ext4_ext_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
3344                                   loff_t offset)
3345 {
3346         struct file *file = iocb->ki_filp;
3347         struct inode *inode = file->f_mapping->host;
3348         ssize_t ret;
3349         size_t count = iov_iter_count(iter);
3350         int overwrite = 0;
3351         get_block_t *get_block_func = NULL;
3352         int dio_flags = 0;
3353         loff_t final_size = offset + count;
3354
3355         /* Use the old path for reads and writes beyond i_size. */
3356         if (iov_iter_rw(iter) != WRITE || final_size > inode->i_size)
3357                 return ext4_ind_direct_IO(iocb, iter, offset);
3358
3359         BUG_ON(iocb->private == NULL);
3360
3361         /*
3362          * Make all waiters for direct IO properly wait also for extent
3363          * conversion. This also disallows race between truncate() and
3364          * overwrite DIO as i_dio_count needs to be incremented under i_mutex.
3365          */
3366         if (iov_iter_rw(iter) == WRITE)
3367                 inode_dio_begin(inode);
3368
3369         /* If we do a overwrite dio, i_mutex locking can be released */
3370         overwrite = *((int *)iocb->private);
3371
3372         if (overwrite)
3373                 inode_unlock(inode);
3374
3375         /*
3376          * We could direct write to holes and fallocate.
3377          *
3378          * Allocated blocks to fill the hole are marked as unwritten to prevent
3379          * parallel buffered read to expose the stale data before DIO complete
3380          * the data IO.
3381          *
3382          * As to previously fallocated extents, ext4 get_block will just simply
3383          * mark the buffer mapped but still keep the extents unwritten.
3384          *
3385          * For non AIO case, we will convert those unwritten extents to written
3386          * after return back from blockdev_direct_IO. That way we save us from
3387          * allocating io_end structure and also the overhead of offloading
3388          * the extent convertion to a workqueue.
3389          *
3390          * For async DIO, the conversion needs to be deferred when the
3391          * IO is completed. The ext4 end_io callback function will be
3392          * called to take care of the conversion work.  Here for async
3393          * case, we allocate an io_end structure to hook to the iocb.
3394          */
3395         iocb->private = NULL;
3396         if (overwrite)
3397                 get_block_func = ext4_dio_get_block_overwrite;
3398         else if (is_sync_kiocb(iocb)) {
3399                 get_block_func = ext4_dio_get_block_unwritten_sync;
3400                 dio_flags = DIO_LOCKING;
3401         } else {
3402                 get_block_func = ext4_dio_get_block_unwritten_async;
3403                 dio_flags = DIO_LOCKING;
3404         }
3405 #ifdef CONFIG_EXT4_FS_ENCRYPTION
3406         BUG_ON(ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode));
3407 #endif
3408         if (IS_DAX(inode))
3409                 ret = dax_do_io(iocb, inode, iter, offset, get_block_func,
3410                                 ext4_end_io_dio, dio_flags);
3411         else
3412                 ret = __blockdev_direct_IO(iocb, inode,
3413                                            inode->i_sb->s_bdev, iter, offset,
3414                                            get_block_func,
3415                                            ext4_end_io_dio, NULL, dio_flags);
3416
3417         if (ret > 0 && !overwrite && ext4_test_inode_state(inode,
3418                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3419                 int err;
3420                 /*
3421                  * for non AIO case, since the IO is already
3422                  * completed, we could do the conversion right here
3423                  */
3424                 err = ext4_convert_unwritten_extents(NULL, inode,
3425                                                      offset, ret);
3426                 if (err < 0)
3427                         ret = err;
3428                 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3429         }
3430
3431         if (iov_iter_rw(iter) == WRITE)
3432                 inode_dio_end(inode);
3433         /* take i_mutex locking again if we do a ovewrite dio */
3434         if (overwrite)
3435                 inode_lock(inode);
3436
3437         return ret;
3438 }
3439
3440 static ssize_t ext4_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
3441                               loff_t offset)
3442 {
3443         struct file *file = iocb->ki_filp;
3444         struct inode *inode = file->f_mapping->host;
3445         size_t count = iov_iter_count(iter);
3446         ssize_t ret;
3447
3448 #ifdef CONFIG_EXT4_FS_ENCRYPTION
3449         if (ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode))
3450                 return 0;
3451 #endif
3452
3453         /*
3454          * If we are doing data journalling we don't support O_DIRECT
3455          */
3456         if (ext4_should_journal_data(inode))
3457                 return 0;
3458
3459         /* Let buffer I/O handle the inline data case. */
3460         if (ext4_has_inline_data(inode))
3461                 return 0;
3462
3463         trace_ext4_direct_IO_enter(inode, offset, count, iov_iter_rw(iter));
3464         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3465                 ret = ext4_ext_direct_IO(iocb, iter, offset);
3466         else
3467                 ret = ext4_ind_direct_IO(iocb, iter, offset);
3468         trace_ext4_direct_IO_exit(inode, offset, count, iov_iter_rw(iter), ret);
3469         return ret;
3470 }
3471
3472 /*
3473  * Pages can be marked dirty completely asynchronously from ext4's journalling
3474  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3475  * much here because ->set_page_dirty is called under VFS locks.  The page is
3476  * not necessarily locked.
3477  *
3478  * We cannot just dirty the page and leave attached buffers clean, because the
3479  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3480  * or jbddirty because all the journalling code will explode.
3481  *
3482  * So what we do is to mark the page "pending dirty" and next time writepage
3483  * is called, propagate that into the buffers appropriately.
3484  */
3485 static int ext4_journalled_set_page_dirty(struct page *page)
3486 {
3487         SetPageChecked(page);
3488         return __set_page_dirty_nobuffers(page);
3489 }
3490
3491 static const struct address_space_operations ext4_aops = {
3492         .readpage               = ext4_readpage,
3493         .readpages              = ext4_readpages,
3494         .writepage              = ext4_writepage,
3495         .writepages             = ext4_writepages,
3496         .write_begin            = ext4_write_begin,
3497         .write_end              = ext4_write_end,
3498         .bmap                   = ext4_bmap,
3499         .invalidatepage         = ext4_invalidatepage,
3500         .releasepage            = ext4_releasepage,
3501         .direct_IO              = ext4_direct_IO,
3502         .migratepage            = buffer_migrate_page,
3503         .is_partially_uptodate  = block_is_partially_uptodate,
3504         .error_remove_page      = generic_error_remove_page,
3505 };
3506
3507 static const struct address_space_operations ext4_journalled_aops = {
3508         .readpage               = ext4_readpage,
3509         .readpages              = ext4_readpages,
3510         .writepage              = ext4_writepage,
3511         .writepages             = ext4_writepages,
3512         .write_begin            = ext4_write_begin,
3513         .write_end              = ext4_journalled_write_end,
3514         .set_page_dirty         = ext4_journalled_set_page_dirty,
3515         .bmap                   = ext4_bmap,
3516         .invalidatepage         = ext4_journalled_invalidatepage,
3517         .releasepage            = ext4_releasepage,
3518         .direct_IO              = ext4_direct_IO,
3519         .is_partially_uptodate  = block_is_partially_uptodate,
3520         .error_remove_page      = generic_error_remove_page,
3521 };
3522
3523 static const struct address_space_operations ext4_da_aops = {
3524         .readpage               = ext4_readpage,
3525         .readpages              = ext4_readpages,
3526         .writepage              = ext4_writepage,
3527         .writepages             = ext4_writepages,
3528         .write_begin            = ext4_da_write_begin,
3529         .write_end              = ext4_da_write_end,
3530         .bmap                   = ext4_bmap,
3531         .invalidatepage         = ext4_da_invalidatepage,
3532         .releasepage            = ext4_releasepage,
3533         .direct_IO              = ext4_direct_IO,
3534         .migratepage            = buffer_migrate_page,
3535         .is_partially_uptodate  = block_is_partially_uptodate,
3536         .error_remove_page      = generic_error_remove_page,
3537 };
3538
3539 void ext4_set_aops(struct inode *inode)
3540 {
3541         switch (ext4_inode_journal_mode(inode)) {
3542         case EXT4_INODE_ORDERED_DATA_MODE:
3543                 ext4_set_inode_state(inode, EXT4_STATE_ORDERED_MODE);
3544                 break;
3545         case EXT4_INODE_WRITEBACK_DATA_MODE:
3546                 ext4_clear_inode_state(inode, EXT4_STATE_ORDERED_MODE);
3547                 break;
3548         case EXT4_INODE_JOURNAL_DATA_MODE:
3549                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3550                 return;
3551         default:
3552                 BUG();
3553         }
3554         if (test_opt(inode->i_sb, DELALLOC))
3555                 inode->i_mapping->a_ops = &ext4_da_aops;
3556         else
3557                 inode->i_mapping->a_ops = &ext4_aops;
3558 }
3559
3560 static int __ext4_block_zero_page_range(handle_t *handle,
3561                 struct address_space *mapping, loff_t from, loff_t length)
3562 {
3563         ext4_fsblk_t index = from >> PAGE_SHIFT;
3564         unsigned offset = from & (PAGE_SIZE-1);
3565         unsigned blocksize, pos;
3566         ext4_lblk_t iblock;
3567         struct inode *inode = mapping->host;
3568         struct buffer_head *bh;
3569         struct page *page;
3570         int err = 0;
3571
3572         page = find_or_create_page(mapping, from >> PAGE_SHIFT,
3573                                    mapping_gfp_constraint(mapping, ~__GFP_FS));
3574         if (!page)
3575                 return -ENOMEM;
3576
3577         blocksize = inode->i_sb->s_blocksize;
3578
3579         iblock = index << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
3580
3581         if (!page_has_buffers(page))
3582                 create_empty_buffers(page, blocksize, 0);
3583
3584         /* Find the buffer that contains "offset" */
3585         bh = page_buffers(page);
3586         pos = blocksize;
3587         while (offset >= pos) {
3588                 bh = bh->b_this_page;
3589                 iblock++;
3590                 pos += blocksize;
3591         }
3592         if (buffer_freed(bh)) {
3593                 BUFFER_TRACE(bh, "freed: skip");
3594                 goto unlock;
3595         }
3596         if (!buffer_mapped(bh)) {
3597                 BUFFER_TRACE(bh, "unmapped");
3598                 ext4_get_block(inode, iblock, bh, 0);
3599                 /* unmapped? It's a hole - nothing to do */
3600                 if (!buffer_mapped(bh)) {
3601                         BUFFER_TRACE(bh, "still unmapped");
3602                         goto unlock;
3603                 }
3604         }
3605
3606         /* Ok, it's mapped. Make sure it's up-to-date */
3607         if (PageUptodate(page))
3608                 set_buffer_uptodate(bh);
3609
3610         if (!buffer_uptodate(bh)) {
3611                 err = -EIO;
3612                 ll_rw_block(READ, 1, &bh);
3613                 wait_on_buffer(bh);
3614                 /* Uhhuh. Read error. Complain and punt. */
3615                 if (!buffer_uptodate(bh))
3616                         goto unlock;
3617                 if (S_ISREG(inode->i_mode) &&
3618                     ext4_encrypted_inode(inode)) {
3619                         /* We expect the key to be set. */
3620                         BUG_ON(!ext4_has_encryption_key(inode));
3621                         BUG_ON(blocksize != PAGE_SIZE);
3622                         WARN_ON_ONCE(ext4_decrypt(page));
3623                 }
3624         }
3625         if (ext4_should_journal_data(inode)) {
3626                 BUFFER_TRACE(bh, "get write access");
3627                 err = ext4_journal_get_write_access(handle, bh);
3628                 if (err)
3629                         goto unlock;
3630         }
3631         zero_user(page, offset, length);
3632         BUFFER_TRACE(bh, "zeroed end of block");
3633
3634         if (ext4_should_journal_data(inode)) {
3635                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3636         } else {
3637                 err = 0;
3638                 mark_buffer_dirty(bh);
3639